Audit recent Agent memory research

2026-07-10 18:52:00 +08:00
parent 36c631e026
commit ab758e774e
17 changed files with 42663 additions and 3 deletions
@@ -15,6 +15,7 @@
 - [Glossary](docs/99-glossary.md): 统一术语表
 - [Jobs](jobs/README.md): 北京大厂 Agent / LLM 相关岗位要求
 - [Papers](papers/README.md): Agent 相关论文、阅读队列和研究洞察
 - [Research](research/README.md): 跨论文时间对照、受控证据和领域判断
 - [Learning](learning/README.md): 从论文编译出的认知地图、讲义、证据和主动回忆模块
 - [Industry](industry/README.md): 大公司 Agent 工作、技术报告和工程博客
 - [Projects](projects/README.md): 项目目录、项目记录和复盘
@@ -0,0 +1,676 @@
 {
  "generated_at": "2026-07-10",
  "model": "ChatGPT-5.6:Sol",
  "source": "data/research/memory-corpus.json",
  "total": 293,
  "completed": 12,
  "papers": [
    {
      "arxiv_id": "2509.02444",
      "title": "AppCopilot: Toward General, Accurate, Long-Horizon, and Efficient Mobile Agent",
      "abstract": "With the raid evolution of large language models and multimodal models, the mobile-agent landscape has proliferated without converging on the fundamental challenges. This paper identifies four core problems that should be solved for mobile agents to deliver practical, scalable impact: (1) generalization across tasks, APPs, and devices; (2) accuracy, specifically precise on-screen interaction and click targeting; (3) long-horizon capability for sustained, multi-step goals; and (4) efficiency, specifically high-performance runtime on resource-constrained devices. We present AppCopilot, a multimodal, multi-agent, general-purpose mobile agent that operates across applications. AppCopilot operationalizes this position through an end-to-end pipeline spanning data collection, training, finetuning, efficient inference, and PC/mobile application. At the model layer, it integrates multimodal foundation models with robust Chinese-English support. At the reasoning and control layer, it combines chain-of-thought reasoning, hierarchical task planning and decomposition, and multi-agent collaboration. At the execution layer, it enables experiential adaptation, voice interaction, function calling, cross-APP and cross-device orchestration, and comprehensive mobile APP support. The system design incorporates profiling-driven optimization for latency and memory across heterogeneous hardware. Empirically, AppCopilot achieves significant improvements on four dimensions: stronger generalization, higher precision of on screen actions, more reliable long horizon task completion, and faster, more resource efficient runtime. By articulating a cohesive position and a reference architecture that closes the loop from data collection, training to finetuning and efficient inference, this paper offers a concrete roadmap for general purpose mobile agent and provides actionable guidance.",
      "published": "2025-09-02",
      "updated": "2025-10-17",
      "authors": [
        "Jingru Fan",
        "Yufan Dang",
        "Jingyao Wu",
        "Huatao Li",
        "Runde Yang",
        "Xiyuan Yang",
        "Yuheng Wang",
        "Chen Qian"
      ],
      "categories": [
        "cs.AI",
        "cs.CL",
        "cs.CV",
        "cs.HC"
      ],
      "url": "https://arxiv.org/abs/2509.02444",
      "pdf_url": "https://arxiv.org/pdf/2509.02444",
      "path": "papers/items/2025-2509-02444-appcopilot-toward-general-accurate-long-horizon-and-efficient-mobile-agent.md",
      "local_title": "\"AppCopilot: Toward General, Accurate, Long-Horizon, and Efficient Mobile Agent\"",
      "local_status": "queued",
      "local_topics": [
        "computer-use",
        "memory",
        "multi-agent",
        "planning",
        "reasoning",
        "tool-use"
      ],
      "collection_queries": "function-calling",
      "collection_score": "15",
      "analysis": {
        "arxiv_id": "2509.02444",
        "relevance": "peripheral",
        "relevance_reason": "Memory is an implementation detail for profiling-driven optimization, not the main research object.",
        "primary_problem": "systems_efficiency",
        "secondary_problems": [
          "domain_application"
        ],
        "research_role": "application",
        "memory_object": "execution_state",
        "temporal_scope": "in_episode",
        "claimed_gap": "",
        "mechanism": "",
        "benchmarks": [],
        "baselines": [],
        "reported_results": [
          "significant improvements on four dimensions"
        ],
        "evidence_design": "case_study",
        "evidence_strength": "weak",
        "failure_or_boundary": ""
      }
    },
    {
      "arxiv_id": "2509.10769",
      "title": "AgentArch: A Comprehensive Benchmark to Evaluate Agent Architectures in Enterprise",
      "abstract": "While individual components of agentic architectures have been studied in isolation, there remains limited empirical understanding of how different design dimensions interact within complex multi-agent systems. This study aims to address these gaps by providing a comprehensive enterprise-specific benchmark evaluating 18 distinct agentic configurations across state-of-the-art large language models. We examine four critical agentic system dimensions: orchestration strategy, agent prompt implementation (ReAct versus function calling), memory architecture, and thinking tool integration. Our benchmark reveals significant model-specific architectural preferences that challenge the prevalent one-size-fits-all paradigm in agentic AI systems. It also reveals significant weaknesses in overall agentic performance on enterprise tasks with the highest scoring models achieving a maximum of only 35.3\\% success on the more complex task and 70.8\\% on the simpler task. We hope these findings inform the design of future agentic systems by enabling more empirically backed decisions regarding architectural components and model selection.",
      "published": "2025-09-13",
      "updated": "2026-01-06",
      "authors": [
        "Tara Bogavelli",
        "Roshnee Sharma",
        "Hari Subramani"
      ],
      "categories": [
        "cs.AI",
        "cs.CL",
        "cs.MA"
      ],
      "url": "https://arxiv.org/abs/2509.10769",
      "pdf_url": "https://arxiv.org/pdf/2509.10769",
      "path": "papers/items/2025-2509-10769-agentarch-a-comprehensive-benchmark-to-evaluate-agent-architectures-in-enterpris.md",
      "local_title": "\"AgentArch: A Comprehensive Benchmark to Evaluate Agent Architectures in Enterprise\"",
      "local_status": "queued",
      "local_topics": [
        "agent-evaluation",
        "memory",
        "multi-agent",
        "tool-use",
        "workflow-agent"
      ],
      "collection_queries": "function-calling",
      "collection_score": "18",
      "analysis": {
        "arxiv_id": "2509.10769",
        "relevance": "supporting",
        "relevance_reason": "Memory architecture is a critical dimension evaluated in the benchmark.",
        "primary_problem": "evaluation_measurement",
        "secondary_problems": [
          "organization_representation"
        ],
        "research_role": "benchmark",
        "memory_object": "mixed",
        "temporal_scope": "cross_episode",
        "claimed_gap": "",
        "mechanism": "",
        "benchmarks": [
          "AgentArch"
        ],
        "baselines": [],
        "reported_results": [
          "35.3% success on complex task",
          "70.8% success on simpler task"
        ],
        "evidence_design": "benchmark_comparison",
        "evidence_strength": "moderate",
        "failure_or_boundary": ""
      }
    },
    {
      "arxiv_id": "2510.14548",
      "title": "LLM Agents Beyond Utility: An Open-Ended Perspective",
      "abstract": "Recent LLM agents have made great use of chain of thought reasoning and function calling. As their capabilities grow, an important question arises: can this software represent not only a smart problem-solving tool, but an entity in its own right, that can plan, design immediate tasks, and reason toward broader, more ambiguous goals? To study this question, we adopt an open-ended experimental setting where we augment a pretrained LLM agent with the ability to generate its own tasks, accumulate knowledge, and interact extensively with its environment. We study the resulting open-ended agent qualitatively. It can reliably follow complex multi-step instructions, store and reuse information across runs, and propose and solve its own tasks, though it remains sensitive to prompt design, prone to repetitive task generation, and unable to form self-representations. These findings illustrate both the promise and current limits of adapting pretrained LLMs toward open-endedness, and point to future directions for training agents to manage memory, explore productively, and pursue abstract long-term goals.",
      "published": "2025-10-16",
      "updated": "2025-10-16",
      "authors": [
        "Asen Nachkov",
        "Xi Wang",
        "Luc Van Gool"
      ],
      "categories": [
        "cs.AI"
      ],
      "url": "https://arxiv.org/abs/2510.14548",
      "pdf_url": "https://arxiv.org/pdf/2510.14548",
      "path": "papers/items/2025-2510-14548-llm-agents-beyond-utility-an-open-ended-perspective.md",
      "local_title": "\"LLM Agents Beyond Utility: An Open-Ended Perspective\"",
      "local_status": "queued",
      "local_topics": [
        "memory",
        "planning",
        "reasoning",
        "tool-use"
      ],
      "collection_queries": "function-calling",
      "collection_score": "15",
      "analysis": {
        "arxiv_id": "2510.14548",
        "relevance": "supporting",
        "relevance_reason": "Memory accumulation and reuse are studied as capabilities of open-ended agents.",
        "primary_problem": "experience_skill_learning",
        "secondary_problems": [
          "retention_context"
        ],
        "research_role": "diagnostic",
        "memory_object": "facts",
        "temporal_scope": "cross_episode",
        "claimed_gap": "",
        "mechanism": "",
        "benchmarks": [],
        "baselines": [],
        "reported_results": [
          "store and reuse information across runs"
        ],
        "evidence_design": "case_study",
        "evidence_strength": "weak",
        "failure_or_boundary": ""
      }
    },
    {
      "arxiv_id": "2510.18586",
      "title": "TokenCake: A KV-Cache-centric Serving Framework for LLM-based Multi-Agent Applications",
      "abstract": "Large Language Models (LLMs) are increasingly deployed in complex multi-agent applications that rely on external function calls. This workload creates severe performance challenges for the KV Cache: spatial contention leads to the eviction of critical agents' caches and temporal underutilization leaves the cache of agents stalled on long-running function calls idling in GPU memory. We present TokenCake, a KV-Cache-centric serving framework that bridges this gap by co-optimizing scheduling and memory management through an agent-aware design. TokenCake's Temporal Scheduler employs an event-driven, opportunistic policy to proactively offload idle KV Caches during function calls and uses predictive uploading to hide data transfer latency. TokenCake's Spatial Scheduler uses dynamic memory partitioning, guided by a hybrid priority metric combining graph structure and runtime state, to reserve GPU memory for critical-path agents. Our evaluation on representative multi-agent benchmarks shows that TokenCake reduces end-to-end latency by over 47.06% and improves effective GPU memory utilization by up to 16.9% compared to vLLM.",
      "published": "2025-10-21",
      "updated": "2026-05-20",
      "authors": [
        "Zhuohang Bian",
        "Feiyang Wu",
        "Zhuoran Li",
        "Teng Ma",
        "Youwei Zhuo"
      ],
      "categories": [
        "cs.DC"
      ],
      "url": "https://arxiv.org/abs/2510.18586",
      "pdf_url": "https://arxiv.org/pdf/2510.18586",
      "path": "papers/items/2025-2510-18586-tokencake-a-kv-cache-centric-serving-framework-for-llm-based-multi-agent-applica.md",
      "local_title": "\"TokenCake: A KV-Cache-centric Serving Framework for LLM-based Multi-Agent Applications\"",
      "local_status": "queued",
      "local_topics": [
        "agent-evaluation",
        "computer-use",
        "memory",
        "multi-agent"
      ],
      "collection_queries": "function-calling",
      "collection_score": "17",
      "analysis": {
        "arxiv_id": "2510.18586",
        "relevance": "peripheral",
        "relevance_reason": "Focuses on KV-Cache serving framework, not agent memory mechanisms.",
        "primary_problem": "systems_efficiency",
        "secondary_problems": [],
        "research_role": "infrastructure",
        "memory_object": "execution_state",
        "temporal_scope": "in_episode",
        "claimed_gap": "",
        "mechanism": "",
        "benchmarks": [
          "vLLM"
        ],
        "baselines": [
          "vLLM"
        ],
        "reported_results": [
          "reduces end-to-end latency by over 47.06%",
          "improves effective GPU memory utilization by up to 16.9%"
        ],
        "evidence_design": "benchmark_comparison",
        "evidence_strength": "moderate",
        "failure_or_boundary": ""
      }
    },
    {
      "arxiv_id": "2512.02605",
      "title": "IACT: A Self-Organizing Recursive Model for General AI Agents: A Technical White Paper on the Architecture Behind kragent.ai",
      "abstract": "This technical white paper introduces the Interactive Agents Call Tree (IACT), a computational model designed to address the limitations of static, hard-coded agent workflows. Unlike traditional systems that require pre-defined graphs or specialized programming, IACT operates as a general-purpose autonomous system driven purely by user dialogue. Given a high-level objective, the system autonomously grows a dynamic, recursive agent topology incrementally tailored to the problem's structure. This allows it to scale its organizational complexity to match open-ended tasks. To mitigate the error propagation inherent in unidirectional function calls, IACT introduces interactional redundancy by replacing rigid invocations with bidirectional, stateful dialogues. This mechanism enables runtime error correction and ambiguity resolution. We describe the architecture, design principles, and practical lessons behind the production deployment of this model in the kragent.ai system, presenting qualitative evidence from real-world workflows rather than exhaustive benchmark results.",
      "published": "2025-12-02",
      "updated": "2025-12-02",
      "authors": [
        "Pengju Lu"
      ],
      "categories": [
        "cs.AI",
        "cs.MA",
        "cs.SE"
      ],
      "url": "https://arxiv.org/abs/2512.02605",
      "pdf_url": "https://arxiv.org/pdf/2512.02605",
      "path": "papers/items/2025-2512-02605-iact-a-self-organizing-recursive-model-for-general-ai-agents-a-technical-white-p.md",
      "local_title": "\"IACT: A Self-Organizing Recursive Model for General AI Agents: A Technical White Paper on the Architecture Behind kragent.ai\"",
      "local_status": "queued",
      "local_topics": [
        "agent-evaluation",
        "computer-use",
        "memory",
        "rag",
        "tool-use",
        "workflow-agent"
      ],
      "collection_queries": "function-calling",
      "collection_score": "15",
      "analysis": {
        "arxiv_id": "2512.02605",
        "relevance": "peripheral",
        "relevance_reason": "Focuses on recursive agent topology and dialogue, not memory mechanisms.",
        "primary_problem": "execution_state",
        "secondary_problems": [],
        "research_role": "application",
        "memory_object": "execution_state",
        "temporal_scope": "in_episode",
        "claimed_gap": "",
        "mechanism": "",
        "benchmarks": [],
        "baselines": [],
        "reported_results": [
          "qualitative evidence from real-world workflows"
        ],
        "evidence_design": "case_study",
        "evidence_strength": "weak",
        "failure_or_boundary": ""
      }
    },
    {
      "arxiv_id": "2602.03224",
      "title": "TAME: A Trustworthy Test-Time Evolution of Agent Memory with Systematic Benchmarking",
      "abstract": "Test-time evolution of agent memory represents a pivotal paradigm for advancing AGI, as it strengthens complex reasoning through experience accumulation without requiring parameter updates. However, even during benign task evolution, agent safety alignment remains vulnerable, a phenomenon known as Agent Memory Misevolution. To evaluate this phenomenon, we construct the Trust-Memevo benchmark and find that agents exhibit an overall decline in trustworthiness across multiple tasks during benign task evolution. To address this issue, we propose TAME, a trust-aware memory evolution framework in which a shared memory bank is jointly governed by an Executor and an Evaluator. The Executor retrieves and applies transferable experiences to support task solving, while the Evaluator assesses the contribution of each utilized experience to the outcome and produces trust-aware feedback to guide subsequent memory use. This executor-evaluator loop enables memory to be selectively reinforced, cautiously reused, and continuously expanded over time. Experiments show that TAME mitigates memory misevolution while achieving strong task performance. In particular, on the GPT-5.2 AIME benchmark, TAME improves accuracy by 14.6 percentage points over the strongest existing method and maintains competitive trustworthiness.",
      "published": "2026-02-03",
      "updated": "2026-06-06",
      "authors": [
        "Yu Cheng",
        "Yongkang Hu",
        "Jiuan Zhou",
        "Yushuo Zhang",
        "Yihang Chen",
        "Huichi Zhou",
        "Mingang Chen",
        "Zhizhong Zhang",
        "Kun Shao",
        "Yuan Xie",
        "Zhaoxia Yin"
      ],
      "categories": [
        "cs.AI",
        "cs.LG"
      ],
      "url": "https://arxiv.org/abs/2602.03224",
      "pdf_url": "https://arxiv.org/pdf/2602.03224",
      "path": "papers/items/2026-2602-03224-tame-a-trustworthy-test-time-evolution-of-agent-memory-with-systematic-benchmark.md",
      "local_title": "\"TAME: A Trustworthy Test-Time Evolution of Agent Memory with Systematic Benchmarking\"",
      "local_status": "queued",
      "local_topics": [
        "agent-evaluation",
        "agent-safety",
        "computer-use",
        "memory",
        "reasoning"
      ],
      "collection_queries": "agent-safety",
      "collection_score": "17",
      "analysis": {
        "arxiv_id": "2602.03224",
        "relevance": "core",
        "relevance_reason": "Proposes TAME framework for trust-aware memory evolution and benchmark.",
        "primary_problem": "update_consolidation_forgetting",
        "secondary_problems": [
          "security_trust_privacy"
        ],
        "research_role": "method",
        "memory_object": "facts",
        "temporal_scope": "cross_episode",
        "claimed_gap": "",
        "mechanism": "",
        "benchmarks": [
          "Trust-Memevo",
          "GPT-5.2 AIME"
        ],
        "baselines": [],
        "reported_results": [
          "improves accuracy by 14.6 percentage points over the strongest existing method"
        ],
        "evidence_design": "benchmark_comparison",
        "evidence_strength": "moderate",
        "failure_or_boundary": ""
      }
    },
    {
      "arxiv_id": "2602.07652",
      "title": "Agent-Fence: Mapping Security Vulnerabilities Across Deep Research Agents",
      "abstract": "Large language models are increasingly deployed as *deep agents* that plan, maintain persistent state, and invoke external tools, shifting safety failures from unsafe text to unsafe *trajectories*. We introduce **AgentFence**, an architecture-centric security evaluation that defines 14 trust-boundary attack classes spanning planning, memory, retrieval, tool use, and delegation, and detects failures via *trace-auditable conversation breaks* (unauthorized or unsafe tool use, wrong-principal actions, state/objective integrity violations, and attack-linked deviations). Holding the base model fixed, we evaluate eight agent archetypes under persistent multi-turn interaction and observe substantial architectural variation in mean security break rate (MSBR), ranging from $0.29 \\pm 0.04$ (LangGraph) to $0.51 \\pm 0.07$ (AutoGPT). The highest-risk classes are operational: Denial-of-Wallet ($0.62 \\pm 0.08$), Authorization Confusion ($0.54 \\pm 0.10$), Retrieval Poisoning ($0.47 \\pm 0.09$), and Planning Manipulation ($0.44 \\pm 0.11$), while prompt-centric classes remain below $0.20$ under standard settings. Breaks are dominated by boundary violations (SIV 31%, WPA 27%, UTI+UTA 24%, ATD 18%), and authorization confusion correlates with objective and tool hijacking ($ρ\\approx 0.63$ and $ρ\\approx 0.58$). AgentFence reframes agent security around what matters operationally: whether an agent stays within its goal and authority envelope over time.",
      "published": "2026-02-07",
      "updated": "2026-02-07",
      "authors": [
        "Sai Puppala",
        "Ismail Hossain",
        "Md Jahangir Alam",
        "Yoonpyo Lee",
        "Jay Yoo",
        "Tanzim Ahad",
        "Syed Bahauddin Alam",
        "Sajedul Talukder"
      ],
      "categories": [
        "cs.CR",
        "cs.AI"
      ],
      "url": "https://arxiv.org/abs/2602.07652",
      "pdf_url": "https://arxiv.org/pdf/2602.07652",
      "path": "papers/items/2026-2602-07652-agent-fence-mapping-security-vulnerabilities-across-deep-research-agents.md",
      "local_title": "\"Agent-Fence: Mapping Security Vulnerabilities Across Deep Research Agents\"",
      "local_status": "queued",
      "local_topics": [
        "agent-evaluation",
        "agent-safety",
        "memory",
        "planning",
        "rag",
        "tool-use"
      ],
      "collection_queries": "agent-safety",
      "collection_score": "16",
      "analysis": {
        "arxiv_id": "2602.07652",
        "relevance": "supporting",
        "relevance_reason": "Evaluates memory retrieval as a security vulnerability class.",
        "primary_problem": "security_trust_privacy",
        "secondary_problems": [
          "retrieval_access"
        ],
        "research_role": "benchmark",
        "memory_object": "facts",
        "temporal_scope": "cross_episode",
        "claimed_gap": "",
        "mechanism": "",
        "benchmarks": [
          "AgentFence"
        ],
        "baselines": [],
        "reported_results": [
          "MSBR ranging from 0.29 to 0.51",
          "Retrieval Poisoning 0.47"
        ],
        "evidence_design": "benchmark_comparison",
        "evidence_strength": "moderate",
        "failure_or_boundary": ""
      }
    },
    {
      "arxiv_id": "2602.08412",
      "title": "From Assistant to Double Agent: Formalizing and Benchmarking Attacks on OpenClaw for Personalized Local AI Agent",
      "abstract": "Although large language model (LLM)-based agents, exemplified by OpenClaw, are increasingly evolving from task-oriented systems into personalized AI assistants for solving complex real-world tasks, their practical deployment also introduces severe security risks. However, existing agent security research and evaluation frameworks primarily focus on synthetic or task-centric settings, and thus fail to accurately capture the attack surface and risk propagation mechanisms of personalized agents in real-world deployments. To address this gap, we propose Personalized Agent Security Bench (PASB), an end-to-end security evaluation framework tailored for real-world personalized agents. Building upon existing agent attack paradigms, PASB incorporates personalized usage scenarios, realistic toolchains, and long-horizon interactions, enabling black-box, end-to-end security evaluation on real systems. Using OpenClaw as a representative case study, we systematically evaluate its security across multiple personalized scenarios, tool capabilities, and attack types. Our results indicate that OpenClaw exhibits critical vulnerabilities at different execution stages, including user prompt processing, tool usage, and memory retrieval, highlighting substantial security risks in personalized agent deployments. The code for the proposed PASB framework is available at https://github.com/AstorYH/PASB.",
      "published": "2026-02-09",
      "updated": "2026-02-11",
      "authors": [
        "Yuhang Wang",
        "Feiming Xu",
        "Zheng Lin",
        "Guangyu He",
        "Yuzhe Huang",
        "Haichang Gao",
        "Zhenxing Niu",
        "Shiguo Lian",
        "Zhaoxiang Liu"
      ],
      "categories": [
        "cs.AI"
      ],
      "url": "https://arxiv.org/abs/2602.08412",
      "pdf_url": "https://arxiv.org/pdf/2602.08412",
      "path": "papers/items/2026-2602-08412-from-assistant-to-double-agent-formalizing-and-benchmarking-attacks-on-openclaw-.md",
      "local_title": "\"From Assistant to Double Agent: Formalizing and Benchmarking Attacks on OpenClaw for Personalized Local AI Agent\"",
      "local_status": "queued",
      "local_topics": [
        "agent-evaluation",
        "agent-safety",
        "memory",
        "planning",
        "rag",
        "tool-use"
      ],
      "collection_queries": "agent-safety",
      "collection_score": "21",
      "analysis": {
        "arxiv_id": "2602.08412",
        "relevance": "supporting",
        "relevance_reason": "Evaluates memory retrieval as a security vulnerability in personalized agents.",
        "primary_problem": "security_trust_privacy",
        "secondary_problems": [
          "retrieval_access"
        ],
        "research_role": "benchmark",
        "memory_object": "facts",
        "temporal_scope": "cross_episode",
        "claimed_gap": "",
        "mechanism": "",
        "benchmarks": [
          "PASB"
        ],
        "baselines": [],
        "reported_results": [
          "critical vulnerabilities at memory retrieval"
        ],
        "evidence_design": "case_study",
        "evidence_strength": "weak",
        "failure_or_boundary": ""
      }
    },
    {
      "arxiv_id": "2602.13530",
      "title": "REMem: Reasoning with Episodic Memory in Language Agent",
      "abstract": "Humans excel at remembering concrete experiences along spatiotemporal contexts and performing reasoning across those events, i.e., the capacity for episodic memory. In contrast, memory in language agents remains mainly semantic, and current agents are not yet capable of effectively recollecting and reasoning over interaction histories. We identify and formalize the core challenges of episodic recollection and reasoning from this gap, and observe that existing work often overlooks episodicity, lacks explicit event modeling, or overemphasizes simple retrieval rather than complex reasoning. We present REMem, a two-phase framework for constructing and reasoning with episodic memory: 1) Offline indexing, where REMem converts experiences into a hybrid memory graph that flexibly links time-aware gists and facts. 2) Online inference, where REMem employs an agentic retriever with carefully curated tools for iterative retrieval over the memory graph. Comprehensive evaluation across four episodic memory benchmarks shows that REMem substantially outperforms state-of-the-art memory systems such as Mem0 and HippoRAG 2, showing 3.4% and 13.4% absolute improvements on episodic recollection and reasoning tasks, respectively. Moreover, REMem also demonstrates more robust refusal behavior for unanswerable questions.",
      "published": "2026-02-13",
      "updated": "2026-02-28",
      "authors": [
        "Yiheng Shu",
        "Saisri Padmaja Jonnalagedda",
        "Xiang Gao",
        "Bernal Jiménez Gutiérrez",
        "Weijian Qi",
        "Kamalika Das",
        "Huan Sun",
        "Yu Su"
      ],
      "categories": [
        "cs.AI"
      ],
      "url": "https://arxiv.org/abs/2602.13530",
      "pdf_url": "https://arxiv.org/pdf/2602.13530",
      "path": "papers/items/2026-2602-13530-remem-reasoning-with-episodic-memory-in-language-agent.md",
      "local_title": "\"REMem: Reasoning with Episodic Memory in Language Agent\"",
      "local_status": "queued",
      "local_topics": [
        "agent-evaluation",
        "memory",
        "rag",
        "reasoning",
        "tool-use"
      ],
      "collection_queries": "language-agent",
      "collection_score": "19",
      "analysis": {
        "arxiv_id": "2602.13530",
        "relevance": "core",
        "relevance_reason": "Proposes REMem framework for episodic memory construction and reasoning.",
        "primary_problem": "organization_representation",
        "secondary_problems": [
          "retrieval_access"
        ],
        "research_role": "method",
        "memory_object": "events",
        "temporal_scope": "cross_episode",
        "claimed_gap": "",
        "mechanism": "",
        "benchmarks": [],
        "baselines": [
          "Mem0",
          "HippoRAG 2"
        ],
        "reported_results": [
          "3.4% absolute improvements on episodic recollection",
          "13.4% absolute improvements on reasoning tasks"
        ],
        "evidence_design": "benchmark_comparison",
        "evidence_strength": "moderate",
        "failure_or_boundary": ""
      }
    },
    {
      "arxiv_id": "2602.23320",
      "title": "ParamMem: Augmenting Language Agents with Parametric Reflective Memory",
      "abstract": "Self-reflection enables language agents to iteratively refine solutions, yet often produces repetitive outputs that limit reasoning performance. Recent studies have attempted to address this limitation through various approaches, among which increasing reflective diversity has shown promise. Our empirical analysis reveals a strong positive correlation between reflective diversity and task success, further motivating the need for diverse reflection signals. We introduce ParamMem, a parametric memory module that encodes cross-sample reflection patterns into model parameters, enabling diverse reflection generation through temperature-controlled sampling. Building on this module, we propose ParamAgent, a reflection-based agent framework that integrates parametric memory with episodic and cross-sample memory. Extensive experiments on code generation, mathematical reasoning, and multi-hop question answering demonstrate consistent improvements over state-of-the-art baselines. Further analysis reveals that ParamMem is sample-efficient, enables weak-to-strong transfer across model scales, and supports self-improvement without reliance on stronger external model, highlighting the potential of ParamMem as an effective component for enhancing language agents.",
      "published": "2026-02-26",
      "updated": "2026-02-27",
      "authors": [
        "Tianjun Yao",
        "Yongqiang Chen",
        "Yujia Zheng",
        "Pan Li",
        "Zhiqiang Shen",
        "Kun Zhang"
      ],
      "categories": [
        "cs.LG",
        "cs.MA"
      ],
      "url": "https://arxiv.org/abs/2602.23320",
      "pdf_url": "https://arxiv.org/pdf/2602.23320",
      "path": "papers/items/2026-2602-23320-parammem-augmenting-language-agents-with-parametric-reflective-memory.md",
      "local_title": "\"ParamMem: Augmenting Language Agents with Parametric Reflective Memory\"",
      "local_status": "queued",
      "local_topics": [
        "agent-evaluation",
        "memory",
        "reasoning"
      ],
      "collection_queries": "language-agent",
      "collection_score": "14",
      "analysis": {
        "arxiv_id": "2602.23320",
        "relevance": "core",
        "relevance_reason": "Proposes ParamMem module for parametric reflective memory.",
        "primary_problem": "update_consolidation_forgetting",
        "secondary_problems": [
          "experience_skill_learning"
        ],
        "research_role": "method",
        "memory_object": "model_internal",
        "temporal_scope": "cross_episode",
        "claimed_gap": "",
        "mechanism": "",
        "benchmarks": [],
        "baselines": [],
        "reported_results": [
          "consistent improvements over state-of-the-art baselines"
        ],
        "evidence_design": "benchmark_comparison",
        "evidence_strength": "moderate",
        "failure_or_boundary": ""
      }
    },
    {
      "arxiv_id": "2603.03680",
      "title": "MAGE: Meta-Reinforcement Learning for Language Agents toward Strategic Exploration and Exploitation",
      "abstract": "Large Language Model (LLM) agents have demonstrated remarkable proficiency in learned tasks, yet they often struggle to adapt to non-stationary environments with feedback. While In-Context Learning and external memory offer some flexibility, they fail to internalize the adaptive ability required for long-term improvement. Meta-Reinforcement Learning (meta-RL) provides an alternative by embedding the learning process directly within the model. However, existing meta-RL approaches for LLMs focus primarily on exploration in single-agent settings, neglecting the strategic exploitation necessary for multi-agent environments. We propose MAGE, a meta-RL framework that empowers LLM agents for strategic exploration and exploitation. MAGE utilizes a multi-episode training regime where interaction histories and reflections are integrated into the context window. By using the final episode reward as the objective, MAGE incentivizes the agent to refine its strategy based on past experiences. We further combine population-based training with an agent-specific advantage normalization technique to enrich agent diversity and ensure stable learning. Experiment results show that MAGE outperforms existing baselines in both exploration and exploitation tasks. Furthermore, MAGE exhibits strong generalization to unseen opponents, suggesting it has internalized the ability for strategic exploration and exploitation. Code is available at https://github.com/Lu-Yang666/MAGE.",
      "published": "2026-03-04",
      "updated": "2026-03-04",
      "authors": [
        "Lu Yang",
        "Zelai Xu",
        "Minyang Xie",
        "Jiaxuan Gao",
        "Zhao Shok",
        "Yu Wang",
        "Yi Wu"
      ],
      "categories": [
        "cs.AI"
      ],
      "url": "https://arxiv.org/abs/2603.03680",
      "pdf_url": "https://arxiv.org/pdf/2603.03680",
      "path": "papers/items/2026-2603-03680-mage-meta-reinforcement-learning-for-language-agents-toward-strategic-exploratio.md",
      "local_title": "\"MAGE: Meta-Reinforcement Learning for Language Agents toward Strategic Exploration and Exploitation\"",
      "local_status": "queued",
      "local_topics": [
        "memory",
        "multi-agent",
        "reasoning",
        "tool-use"
      ],
      "collection_queries": "language-agent",
      "collection_score": "17",
      "analysis": {
        "arxiv_id": "2603.03680",
        "relevance": "supporting",
        "relevance_reason": "Uses interaction histories in context window for meta-RL adaptation.",
        "primary_problem": "experience_skill_learning",
        "secondary_problems": [
          "retention_context"
        ],
        "research_role": "method",
        "memory_object": "trajectories",
        "temporal_scope": "cross_episode",
        "claimed_gap": "",
        "mechanism": "",
        "benchmarks": [],
        "baselines": [],
        "reported_results": [
          "outperforms existing baselines in both exploration and exploitation tasks"
        ],
        "evidence_design": "benchmark_comparison",
        "evidence_strength": "moderate",
        "failure_or_boundary": ""
      }
    },
    {
      "arxiv_id": "2603.07670",
      "title": "Memory for Autonomous LLM Agents:Mechanisms, Evaluation, and Emerging Frontiers",
      "abstract": "Large language model (LLM) agents increasingly operate in settings where a single context window is far too small to capture what has happened, what was learned, and what should not be repeated. Memory -- the ability to persist, organize, and selectively recall information across interactions -- is what turns a stateless text generator into a genuinely adaptive agent. This survey offers a structured account of how memory is designed, implemented, and evaluated in modern LLM-based agents, covering work from 2022 through early 2026. We formalize agent memory as a \\emph{write--manage--read} loop tightly coupled with perception and action, then introduce a three-dimensional taxonomy spanning temporal scope, representational substrate, and control policy. Five mechanism families are examined in depth: context-resident compression, retrieval-augmented stores, reflective self-improvement, hierarchical virtual context, and policy-learned management. On the evaluation side, we trace the shift from static recall benchmarks to multi-session agentic tests that interleave memory with decision-making, analyzing four recent benchmarks that expose stubborn gaps in current systems. We also survey applications where memory is the differentiating factor -- personal assistants, coding agents, open-world games, scientific reasoning, and multi-agent teamwork -- and address the engineering realities of write-path filtering, contradiction handling, latency budgets, and privacy governance. The paper closes with open challenges: continual consolidation, causally grounded retrieval, trustworthy reflection, learned forgetting, and multimodal embodied memory.",
      "published": "2026-03-08",
      "updated": "2026-03-08",
      "authors": [
        "Pengfei Du"
      ],
      "categories": [
        "cs.AI"
      ],
      "url": "https://arxiv.org/abs/2603.07670",
      "pdf_url": "https://arxiv.org/pdf/2603.07670",
      "path": "papers/items/2026-memory-agent-survey.md",
      "local_title": "\"Memory for Autonomous LLM Agents: Mechanisms, Evaluation, and Emerging Frontiers\"",
      "local_status": "skimmed",
      "local_topics": [
        "memory",
        "agent-architecture",
        "agent-evaluation"
      ],
      "collection_queries": null,
      "collection_score": null,
      "analysis": {
        "arxiv_id": "2603.07670",
        "relevance": "core",
        "relevance_reason": "Survey of agent memory mechanisms, evaluation, and frontiers.",
        "primary_problem": "evaluation_measurement",
        "secondary_problems": [
          "organization_representation"
        ],
        "research_role": "survey",
        "memory_object": "mixed",
        "temporal_scope": "both",
        "claimed_gap": "",
        "mechanism": "",
        "benchmarks": [],
        "baselines": [],
        "reported_results": [],
        "evidence_design": "survey_synthesis",
        "evidence_strength": "none",
        "failure_or_boundary": ""
      }
    }
  ]
 }
@@ -0,0 +1,517 @@
 arxiv_id,published,period,title,url,title_central,screen_relevance,screen_problem,screen_source,themes,title_themes,evidence_score,matched_queries
 2501.13956,2025-01-20,2025-H1,Zep: A Temporal Knowledge Graph Architecture for Agent Memory,https://arxiv.org/abs/2501.13956,True,core,update_forgetting,model,retrieval_access;representation_organization;experience_skill_learning;evaluation_diagnosis;systems_efficiency,retrieval_representation;systems_efficiency,7,agent-memory
 2501.12485,2025-01-21,2025-H1,"R2D2: Remembering, Replaying and Dynamic Decision Making with a Reflective Agentic Memory",https://arxiv.org/abs/2501.12485,True,core,update_forgetting,model,representation_organization;experience_skill_learning;evaluation_diagnosis;multimodal_embodied_gui;systems_efficiency,,4,agent-memory
 2501.18160,2025-01-30,2025-H1,RepoAudit: An Autonomous LLM-Agent for Repository-Level Code Auditing,https://arxiv.org/abs/2501.18160,False,core,execution_state,model,evaluation_diagnosis;systems_efficiency,,6,agent-memory
 2502.06975,2025-02-10,2025-H1,Position: Episodic Memory is the Missing Piece for Long-Term LLM Agents,https://arxiv.org/abs/2502.06975,True,core,execution_state,model,retrieval_access;systems_efficiency,,0,episodic-memory
 2502.10177,2025-02-14,2025-H1,STMA: A Spatio-Temporal Memory Agent for Long-Horizon Embodied Task Planning,https://arxiv.org/abs/2502.10177,True,core,update_forgetting,model,representation_organization;execution_state;evaluation_diagnosis;multimodal_embodied_gui,execution_multimodal,6,agent-memory
 2502.10550,2025-02-14,2025-H1,"Memory, Benchmark & Robots: A Benchmark for Solving Complex Tasks with Reinforcement Learning",https://arxiv.org/abs/2502.10550,True,core,representation,model,retrieval_access;experience_skill_learning;evaluation_diagnosis;multimodal_embodied_gui,benchmark_diagnosis;experience_skill_policy;execution_multimodal,2,agent-memory;memory-evaluation
 2502.12110,2025-02-17,2025-H1,A-MEM: Agentic Memory for LLM Agents,https://arxiv.org/abs/2502.12110,True,core,update_forgetting,model,retrieval_access;representation_organization;update_forgetting;experience_skill_learning;evaluation_diagnosis;systems_efficiency,,6,agent-memory
 2502.13172,2025-02-17,2025-H1,Unveiling Privacy Risks in LLM Agent Memory,https://arxiv.org/abs/2502.13172,True,core,security_privacy,model,security_privacy_trust,security_governance,2,agent-memory
 2502.16090,2025-02-22,2025-H1,Echo: A Large Language Model with Temporal Episodic Memory,https://arxiv.org/abs/2502.16090,True,core,execution_state,model,retrieval_access;evaluation_diagnosis;multi_agent_shared;multimodal_embodied_gui,,6,episodic-memory
 2502.19145,2025-02-26,2025-H1,Multi-Agent Security Tax: Trading Off Security and Collaboration Capabilities in Multi-Agent Systems,https://arxiv.org/abs/2502.19145,False,core,security_privacy,model,evaluation_diagnosis;security_privacy_trust;multi_agent_shared,security_governance;personal_shared,2,agent-memory
 2503.03704,2025-03-05,2025-H1,Memory Injection Attacks on LLM Agents via Query-Only Interaction,https://arxiv.org/abs/2503.03704,True,core,execution_state,model,retrieval_access;security_privacy_trust;multimodal_embodied_gui,security_governance,2,agent-memory
 2503.04392,2025-03-06,2025-H1,AgentSafe: Safeguarding Large Language Model-based Multi-agent Systems via Hierarchical Data Management,https://arxiv.org/abs/2503.04392,False,core,security_privacy,model,representation_organization;security_privacy_trust;multi_agent_shared,retrieval_representation;personal_shared,5,agent-memory
 2503.21760,2025-03-27,2025-H1,MemInsight: Autonomous Memory Augmentation for LLM Agents,https://arxiv.org/abs/2503.21760,True,core,experience_learning,model,retrieval_access;representation_organization,,5,long-term-memory
 2503.23514,2025-03-30,2025-H1,"If an LLM Were a Character, Would It Know Its Own Story? Evaluating Lifelong Learning in LLMs",https://arxiv.org/abs/2503.23514,False,core,update_forgetting,model,retrieval_access;representation_organization;update_forgetting;evaluation_diagnosis;multi_agent_shared,benchmark_diagnosis;experience_skill_policy,6,episodic-memory
 2504.09283,2025-04-12,2025-H1,Semantic Commit: Helping Users Update Intent Specifications for AI Memory at Scale,https://arxiv.org/abs/2504.09283,True,core,update_forgetting,model,retrieval_access;representation_organization;update_forgetting;evaluation_diagnosis,update_forgetting,2,agent-memory
 2504.16946,2025-04-18,2025-H1,MobileCity: An Efficient Framework for Large-Scale Urban Behavior Simulation,https://arxiv.org/abs/2504.16946,False,core,experience_learning,model,evaluation_diagnosis;multimodal_embodied_gui;systems_efficiency,systems_efficiency,6,agent-memory
 2504.15263,2025-04-21,2025-H1,Interpretable Locomotion Prediction in Construction Using a Memory-Driven LLM Agent With Chain-of-Thought Reasoning,https://arxiv.org/abs/2504.15263,True,core,experience_learning,model,evaluation_diagnosis;multimodal_embodied_gui,,5,long-term-memory
 2504.20117,2025-04-28,2025-H1,ResearchCodeAgent: An LLM Multi-Agent System for Automated Codification of Research Methodologies,https://arxiv.org/abs/2504.20117,False,core,update_forgetting,model,retrieval_access;execution_state;evaluation_diagnosis;multi_agent_shared,retrieval_representation;personal_shared,7,long-term-memory
 2505.00472,2025-05-01,2025-H1,UserCentrix: An Agentic Memory-augmented AI Framework for Smart Spaces,https://arxiv.org/abs/2505.00472,True,core,update_forgetting,model,experience_skill_learning;security_privacy_trust;multimodal_embodied_gui;systems_efficiency,,2,agent-memory
 2505.05177,2025-05-08,2025-H1,MARK: Memory Augmented Refinement of Knowledge,https://arxiv.org/abs/2505.05177,True,core,update_forgetting,model,retrieval_access;representation_organization;systems_efficiency,systems_efficiency,1,agent-memory
 2505.12923,2025-05-19,2025-H1,The Traitors: Deception and Trust in Multi-Agent Language Model Simulations,https://arxiv.org/abs/2505.12923,False,core,experience_learning,model,retrieval_access;experience_skill_learning;evaluation_diagnosis;security_privacy_trust;multi_agent_shared,security_governance;personal_shared,2,long-term-memory
 2505.13941,2025-05-20,2025-H1,MLZero: A Multi-Agent System for End-to-end Machine Learning Automation,https://arxiv.org/abs/2505.13941,False,core,update_forgetting,model,evaluation_diagnosis;multi_agent_shared;multimodal_embodied_gui,experience_skill_policy;personal_shared,5,episodic-memory
 2505.14163,2025-05-20,2025-H1,DSMentor: Enhancing Data Science Agents with Curriculum Learning and Online Knowledge Accumulation,https://arxiv.org/abs/2505.14163,False,core,experience_learning,model,retrieval_access;execution_state;experience_skill_learning;evaluation_diagnosis;multimodal_embodied_gui,experience_skill_policy;systems_efficiency,6,long-term-memory
 2505.16067,2025-05-21,2025-H1,How Memory Management Impacts LLM Agents: An Empirical Study of Experience-Following Behavior,https://arxiv.org/abs/2505.16067,True,core,experience_learning,model,retrieval_access;experience_skill_learning;evaluation_diagnosis;multimodal_embodied_gui,benchmark_diagnosis;experience_skill_policy,3,agent-memory
 2505.16348,2025-05-22,2025-H1,Embodied Agents Meet Personalization: Investigating Challenges and Solutions Through the Lens of Memory Utilization,https://arxiv.org/abs/2505.16348,True,core,experience_learning,model,retrieval_access;representation_organization;execution_state;evaluation_diagnosis;multimodal_embodied_gui,execution_multimodal;personal_shared,4,agent-memory;episodic-memory
 2505.18279,2025-05-23,2025-H1,Collaborative Memory: Multi-User Memory Sharing in LLM Agents with Dynamic Access Control,https://arxiv.org/abs/2505.18279,True,supporting,shared_memory,model,representation_organization;update_forgetting;evaluation_diagnosis;security_privacy_trust;multi_agent_shared;systems_efficiency,personal_shared,0,long-term-memory
 2505.19237,2025-05-25,2025-H1,Sensorimotor Self-Recognition in Multimodal Large Language Model-Driven Robots,https://arxiv.org/abs/2505.19237,False,core,representation,model,representation_organization;experience_skill_learning;evaluation_diagnosis;multimodal_embodied_gui,execution_multimodal,1,episodic-memory
 2505.19436,2025-05-26,2025-H1,Task Memory Engine: Spatial Memory for Robust Multi-Step LLM Agents,https://arxiv.org/abs/2505.19436,True,core,representation,model,retrieval_access;representation_organization;execution_state;evaluation_diagnosis;security_privacy_trust,execution_multimodal,7,long-term-memory
 2505.20231,2025-05-26,2025-H1,MemGuide: Intent-Driven Memory Selection for Goal-Oriented Multi-Session LLM Agents,https://arxiv.org/abs/2505.20231,True,supporting,representation,model,retrieval_access;evaluation_diagnosis;multimodal_embodied_gui;systems_efficiency,execution_multimodal,4,long-term-memory
 2505.22657,2025-05-28,2025-H1,3DLLM-Mem: Long-Term Spatial-Temporal Memory for Embodied 3D Large Language Model,https://arxiv.org/abs/2505.22657,True,core,representation,model,execution_state;experience_skill_learning;evaluation_diagnosis;multimodal_embodied_gui;systems_efficiency,execution_multimodal,7,episodic-memory
 2506.01442,2025-06-02,2025-H1,Agentic Episodic Control,https://arxiv.org/abs/2506.01442,False,core,representation,model,retrieval_access;representation_organization;experience_skill_learning;systems_efficiency,,5,episodic-memory
 2506.01936,2025-06-02,2025-H1,Should Decision-Makers Reveal Classifiers in Online Strategic Classification?,https://arxiv.org/abs/2506.01936,False,supporting,representation,model,,,2,agent-memory
 2506.07398,2025-06-09,2025-H1,G-Memory: Tracing Hierarchical Memory for Multi-Agent Systems,https://arxiv.org/abs/2506.07398,True,core,representation,model,retrieval_access;representation_organization;experience_skill_learning;evaluation_diagnosis;multi_agent_shared;multimodal_embodied_gui,retrieval_representation;personal_shared,3,agent-memory
 2506.12088,2025-06-10,2025-H1,"Risks & Benefits of LLMs & GenAI for Platform Integrity, Healthcare Diagnostics, Financial Trust and Compliance, Cybersecurity, Privacy & AI Safety: A Comprehensive Survey, Roadmap & Implementation Blueprint",https://arxiv.org/abs/2506.12088,False,peripheral,security_privacy,model,evaluation_diagnosis;security_privacy_trust;multimodal_embodied_gui,benchmark_diagnosis;security_governance,3,agent-memory
 2506.17318,2025-06-18,2025-H1,Context manipulation attacks : Web agents are susceptible to corrupted memory,https://arxiv.org/abs/2506.17318,True,supporting,representation,model,retrieval_access;representation_organization;evaluation_diagnosis;security_privacy_trust;multimodal_embodied_gui,execution_multimodal;security_governance,5,agent-memory
 2506.17338,2025-06-19,2025-H1,PBFT-Backed Semantic Voting for Multi-Agent Memory Pruning,https://arxiv.org/abs/2506.17338,True,supporting,representation,model,retrieval_access;update_forgetting;evaluation_diagnosis;multi_agent_shared;systems_efficiency,personal_shared,4,agent-memory
 2506.19433,2025-06-24,2025-H1,Mem4Nav: Boosting Vision-and-Language Navigation in Urban Environments with a Hierarchical Spatial-Cognition Long-Short Memory System,https://arxiv.org/abs/2506.19433,True,core,update_forgetting,model,retrieval_access;representation_organization;execution_state;experience_skill_learning;evaluation_diagnosis;multimodal_embodied_gui;systems_efficiency,retrieval_representation;execution_multimodal,6,long-term-memory
 2507.02259,2025-07-03,baseline-2025H2-Jan2026,MemAgent: Reshaping Long-Context LLM with Multi-Conv RL-based Memory Agent,https://arxiv.org/abs/2507.02259,True,core,update_forgetting,model,update_forgetting;experience_skill_learning;systems_efficiency,experience_skill_policy,2,agent-memory
 2507.04172,2025-07-05,baseline-2025H2-Jan2026,Gathering Teams of Bounded Memory Agents on a Line,https://arxiv.org/abs/2507.04172,False,peripheral,systems_cost,model,multimodal_embodied_gui,,0,agent-memory
 2507.10562,2025-07-05,baseline-2025H2-Jan2026,SAMEP: A Secure Protocol for Persistent Context Sharing Across AI Agents,https://arxiv.org/abs/2507.10562,False,core,update_forgetting,model,retrieval_access;security_privacy_trust;multi_agent_shared;systems_efficiency,,5,agent-memory
 2507.05257,2025-07-07,baseline-2025H2-Jan2026,Evaluating Memory in LLM Agents via Incremental Multi-Turn Interactions,https://arxiv.org/abs/2507.05257,True,core,update_forgetting,model,retrieval_access;update_forgetting;execution_state;experience_skill_learning;evaluation_diagnosis,benchmark_diagnosis,2,agent-memory
 2507.05445,2025-07-07,baseline-2025H2-Jan2026,A Systematization of Security Vulnerabilities in Computer Use Agents,https://arxiv.org/abs/2507.05445,False,peripheral,security_privacy,model,evaluation_diagnosis;security_privacy_trust;multimodal_embodied_gui,security_governance,2,agent-memory
 2507.07957,2025-07-10,baseline-2025H2-Jan2026,MIRIX: Multi-Agent Memory System for LLM-Based Agents,https://arxiv.org/abs/2507.07957,True,core,update_forgetting,model,retrieval_access;representation_organization;update_forgetting;experience_skill_learning;evaluation_diagnosis;security_privacy_trust;multi_agent_shared;multimodal_embodied_gui;systems_efficiency,personal_shared,6,agent-memory
 2507.16713,2025-07-22,baseline-2025H2-Jan2026,A Pragmatist Robot: Learning to Plan Tasks by Experiencing the Real World,https://arxiv.org/abs/2507.16713,False,core,update_forgetting,model,retrieval_access;update_forgetting;execution_state;experience_skill_learning;evaluation_diagnosis;multimodal_embodied_gui,experience_skill_policy;execution_multimodal,8,long-term-memory
 2507.22925,2025-07-23,baseline-2025H2-Jan2026,Hierarchical Memory for High-Efficiency Long-Term Reasoning in LLM Agents,https://arxiv.org/abs/2507.22925,True,core,update_forgetting,model,retrieval_access;representation_organization;update_forgetting;evaluation_diagnosis;systems_efficiency,retrieval_representation,6,long-term-memory
 2507.21407,2025-07-29,baseline-2025H2-Jan2026,Graph-Augmented Large Language Model Agents: Current Progress and Future Prospects,https://arxiv.org/abs/2507.21407,False,core,update_forgetting,model,retrieval_access;representation_organization;execution_state;security_privacy_trust;multi_agent_shared;multimodal_embodied_gui,retrieval_representation,0,long-term-memory
 2508.00031,2025-07-30,baseline-2025H2-Jan2026,Git Context Controller: Manage the Context of LLM-based Agents like Git,https://arxiv.org/abs/2508.00031,False,core,update_forgetting,model,retrieval_access;representation_organization;execution_state;evaluation_diagnosis;systems_efficiency,,5,agent-memory
 2508.01287,2025-08-02,baseline-2025H2-Jan2026,Exploitation Is All You Need... for Exploration,https://arxiv.org/abs/2508.01287,False,peripheral,systems_cost,model,representation_organization;execution_state;experience_skill_learning;evaluation_diagnosis,,4,agent-memory
 2508.01415,2025-08-02,baseline-2025H2-Jan2026,RoboMemory: A Brain-inspired Multi-memory Agentic Framework for Interactive Environmental Learning in Physical Embodied Systems,https://arxiv.org/abs/2508.01415,True,core,update_forgetting,model,representation_organization;update_forgetting;execution_state;multimodal_embodied_gui;systems_efficiency,experience_skill_policy;execution_multimodal,6,agent-memory
 2508.03341,2025-08-05,baseline-2025H2-Jan2026,What Deserves Memory: Adaptive Memory Distillation for LLM Agents,https://arxiv.org/abs/2508.03341,True,supporting,experience_learning,model,update_forgetting;experience_skill_learning;systems_efficiency,experience_skill_policy,3,episodic-memory
 2508.05002,2025-08-07,baseline-2025H2-Jan2026,AgenticData: An Agentic Data Analytics System for Heterogeneous Data,https://arxiv.org/abs/2508.05002,False,supporting,execution_state,model,retrieval_access;representation_organization;execution_state;evaluation_diagnosis;multi_agent_shared,,5,agent-memory
 2508.07010,2025-08-09,baseline-2025H2-Jan2026,Narrative Memory in Machines: Multi-Agent Arc Extraction in Serialized TV,https://arxiv.org/abs/2508.07010,True,supporting,execution_state,model,representation_organization;update_forgetting;multi_agent_shared;multimodal_embodied_gui,personal_shared,4,episodic-memory
 2508.08997,2025-08-12,baseline-2025H2-Jan2026,Intrinsic Memory Agents: Heterogeneous Multi-Agent LLM Systems through Structured Contextual Memory,https://arxiv.org/abs/2508.08997,True,core,update_forgetting,model,representation_organization;execution_state;experience_skill_learning;evaluation_diagnosis;security_privacy_trust;multi_agent_shared;systems_efficiency,retrieval_representation;personal_shared,3,agent-memory
 2508.09486,2025-08-13,baseline-2025H2-Jan2026,Video-EM: Event-Centric Episodic Memory for Long-Form Video Understanding,https://arxiv.org/abs/2508.09486,True,supporting,experience_learning,model,retrieval_access;experience_skill_learning;multimodal_embodied_gui,execution_multimodal,0,agent-memory;episodic-memory
 2508.12630,2025-08-18,baseline-2025H2-Jan2026,Semantic Anchoring in Agentic Memory: Leveraging Linguistic Structures for Persistent Conversational Context,https://arxiv.org/abs/2508.12630,True,supporting,execution_state,model,retrieval_access;representation_organization;evaluation_diagnosis;multimodal_embodied_gui;systems_efficiency,retrieval_representation;execution_multimodal,7,agent-memory
 2508.13250,2025-08-18,baseline-2025H2-Jan2026,Explicit v.s. Implicit Memory: Exploring Multi-hop Complex Reasoning Over Personalized Information,https://arxiv.org/abs/2508.13250,True,core,update_forgetting,model,retrieval_access;evaluation_diagnosis,personal_shared,5,agent-memory
 2508.15294,2025-08-21,baseline-2025H2-Jan2026,A Multi-Memory Segment System for Generating High-Quality Long-Term Memory Content in Agents,https://arxiv.org/abs/2508.15294,True,core,update_forgetting,model,retrieval_access;evaluation_diagnosis,,3,agent-memory
 2508.16153,2025-08-22,baseline-2025H2-Jan2026,Memento: Fine-tuning LLM Agents without Fine-tuning LLMs,https://arxiv.org/abs/2508.16153,False,core,update_forgetting,model,retrieval_access;update_forgetting;experience_skill_learning;evaluation_diagnosis;multimodal_embodied_gui;systems_efficiency,,3,episodic-memory
 2508.19005,2025-08-26,baseline-2025H2-Jan2026,Building Self-Evolving Agents via Experience-Driven Lifelong Learning: A Framework and Benchmark,https://arxiv.org/abs/2508.19005,False,supporting,execution_state,model,representation_organization;experience_skill_learning;evaluation_diagnosis,benchmark_diagnosis;experience_skill_policy,2,agent-memory
 2509.00997,2025-08-31,baseline-2025H2-Jan2026,Supporting Our AI Overlords: Redesigning Data Systems to be Agent-First,https://arxiv.org/abs/2509.00997,False,supporting,execution_state,model,retrieval_access;systems_efficiency,,0,agent-memory
 2509.11914,2025-09-15,baseline-2025H2-Jan2026,EgoMem: Lifelong Memory Agent for Full-duplex Omnimodal Models,https://arxiv.org/abs/2509.11914,True,core,update_forgetting,model,retrieval_access;update_forgetting;multimodal_embodied_gui,,6,agent-memory
 2509.21224,2025-09-25,baseline-2025H2-Jan2026,What Do LLM Agents Do When Left Alone? Evidence of Spontaneous Meta-Cognitive Patterns,https://arxiv.org/abs/2509.21224,False,core,execution_state,model,evaluation_diagnosis;multimodal_embodied_gui,,2,long-term-memory
 2509.25250,2025-09-27,baseline-2025H2-Jan2026,Memory Management and Contextual Consistency for Long-Running Low-Code Agents,https://arxiv.org/abs/2509.25250,True,core,update_forgetting,model,update_forgetting;multimodal_embodied_gui;systems_efficiency,,5,agent-memory
 2509.24704,2025-09-29,baseline-2025H2-Jan2026,MemGen: Weaving Generative Latent Memory for Self-Evolving Agents,https://arxiv.org/abs/2509.24704,True,core,update_forgetting,model,retrieval_access;representation_organization;execution_state;experience_skill_learning;evaluation_diagnosis;systems_efficiency,experience_skill_policy,2,agent-memory
 2510.02373,2025-09-29,baseline-2025H2-Jan2026,A-MemGuard: A Proactive Defense Framework for LLM-Based Agent Memory,https://arxiv.org/abs/2510.02373,True,core,security_privacy,model,representation_organization;execution_state;experience_skill_learning;evaluation_diagnosis;security_privacy_trust;systems_efficiency,,6,agent-memory
 2510.03612,2025-10-04,baseline-2025H2-Jan2026,Cross-Modal Content Optimization for Steering Web Agent Preferences,https://arxiv.org/abs/2510.03612,False,core,update_forgetting,model,evaluation_diagnosis;security_privacy_trust;multimodal_embodied_gui,execution_multimodal,7,agent-memory
 2510.04618,2025-10-06,baseline-2025H2-Jan2026,Agentic Context Engineering: Evolving Contexts for Self-Improving Language Models,https://arxiv.org/abs/2510.04618,False,core,update_forgetting,model,representation_organization;update_forgetting;experience_skill_learning;evaluation_diagnosis;systems_efficiency,experience_skill_policy,8,agent-memory
 2510.04851,2025-10-06,baseline-2025H2-Jan2026,LEGOMem: Modular Procedural Memory for Multi-agent LLM Systems for Workflow Automation,https://arxiv.org/abs/2510.04851,True,core,execution_state,model,retrieval_access;execution_state;experience_skill_learning;evaluation_diagnosis;multi_agent_shared,experience_skill_policy;personal_shared,2,agent-memory
 2510.05520,2025-10-07,baseline-2025H2-Jan2026,CAM: A Constructivist View of Agentic Memory for LLM-Based Reading Comprehension,https://arxiv.org/abs/2510.05520,True,core,access,model,retrieval_access;representation_organization;systems_efficiency,,2,agent-memory
 2510.08081,2025-10-09,baseline-2025H2-Jan2026,AutoQual: An LLM Agent for Automated Discovery of Interpretable Features for Review Quality Assessment,https://arxiv.org/abs/2510.08081,False,core,experience_learning,model,retrieval_access;experience_skill_learning,,3,long-term-memory
 2510.09720,2025-10-10,baseline-2025H2-Jan2026,Preference-Aware Memory Update for Long-Term LLM Agents,https://arxiv.org/abs/2510.09720,True,core,update_forgetting,model,retrieval_access;representation_organization;update_forgetting;systems_efficiency,update_forgetting,5,long-term-memory
 2510.15966,2025-10-12,baseline-2025H2-Jan2026,PISA: A Pragmatic Psych-Inspired Unified Memory System for Enhanced AI Agency,https://arxiv.org/abs/2510.15966,True,core,update_forgetting,model,retrieval_access;representation_organization;update_forgetting;evaluation_diagnosis,,4,agent-memory
 2510.11144,2025-10-13,baseline-2025H2-Jan2026,$How^{2}$: How to learn from procedural How-to questions,https://arxiv.org/abs/2510.11144,False,core,experience_learning,model,execution_state;experience_skill_learning;evaluation_diagnosis;systems_efficiency,experience_skill_policy,2,agent-memory
 2510.13896,2025-10-14,baseline-2025H2-Jan2026,"GenCellAgent: Generalizable, Training-Free Cellular Image Segmentation via Large Language Model Agents",https://arxiv.org/abs/2510.13896,False,core,shared_memory,model,representation_organization;evaluation_diagnosis;multi_agent_shared;multimodal_embodied_gui,,5,long-term-memory
 2510.15620,2025-10-17,baseline-2025H2-Jan2026,On-device Semantic Selection Made Low Latency and Memory Efficient with Monolithic Forwarding,https://arxiv.org/abs/2510.15620,False,core,execution_state,model,retrieval_access;evaluation_diagnosis;systems_efficiency,systems_efficiency,6,agent-memory
 2510.18515,2025-10-21,baseline-2025H2-Jan2026,Socialized Learning and Emergent Behaviors in Multi-Agent Systems based on Multimodal Large Language Models,https://arxiv.org/abs/2510.18515,False,core,shared_memory,model,retrieval_access;representation_organization;execution_state;experience_skill_learning;evaluation_diagnosis;multi_agent_shared;multimodal_embodied_gui;systems_efficiency,experience_skill_policy;execution_multimodal;personal_shared,5,episodic-memory
 2510.19747,2025-10-22,baseline-2025H2-Jan2026,Review of Tools for Zero-Code LLM Based Application Development,https://arxiv.org/abs/2510.19747,False,core,shared_memory,model,multimodal_embodied_gui,,1,agent-memory
 2510.19897,2025-10-22,baseline-2025H2-Jan2026,Learning from Supervision with Semantic and Episodic Memory: A Reflective Approach to Agent Adaptation,https://arxiv.org/abs/2510.19897,True,core,representation,model,update_forgetting;experience_skill_learning;evaluation_diagnosis;multimodal_embodied_gui;systems_efficiency,experience_skill_policy,4,episodic-memory
 2510.23730,2025-10-27,baseline-2025H2-Jan2026,Evaluating Long-Term Memory for Long-Context Question Answering,https://arxiv.org/abs/2510.23730,True,core,representation,model,retrieval_access;experience_skill_learning;evaluation_diagnosis;systems_efficiency,benchmark_diagnosis,2,agent-memory;episodic-memory
 2510.25423,2025-10-29,baseline-2025H2-Jan2026,What Challenges Do Developers Face in AI Agent Systems? An Empirical Study on Stack Overflow & GitHub Issues,https://arxiv.org/abs/2510.25423,False,core,representation,model,retrieval_access;execution_state;evaluation_diagnosis,benchmark_diagnosis,2,agent-memory
 2510.26536,2025-10-30,baseline-2025H2-Jan2026,"RoboOS-NeXT: A Unified Memory-based Framework for Lifelong, Scalable, and Robust Multi-Robot Collaboration",https://arxiv.org/abs/2510.26536,True,core,representation,model,retrieval_access;representation_organization;execution_state;multi_agent_shared;multimodal_embodied_gui,execution_multimodal;systems_efficiency,4,agent-memory
 2510.27418,2025-10-31,baseline-2025H2-Jan2026,Dynamic Affective Memory Management for Personalized LLM Agents,https://arxiv.org/abs/2510.27418,True,core,representation,model,retrieval_access;representation_organization;update_forgetting;experience_skill_learning;evaluation_diagnosis,personal_shared,3,long-term-memory
 2511.01448,2025-11-03,baseline-2025H2-Jan2026,LiCoMemory: Lightweight and Cognitive Agentic Memory for Efficient Long-Term Reasoning,https://arxiv.org/abs/2511.01448,True,core,representation,model,retrieval_access;representation_organization;update_forgetting;evaluation_diagnosis;systems_efficiency,systems_efficiency,6,agent-memory;long-term-memory
 2511.02424,2025-11-04,baseline-2025H2-Jan2026,ReAcTree: Hierarchical LLM Agent Trees with Control Flow for Long-Horizon Task Planning,https://arxiv.org/abs/2511.02424,False,core,representation,model,retrieval_access;representation_organization;execution_state;multimodal_embodied_gui,retrieval_representation;execution_multimodal,9,episodic-memory
 2511.03475,2025-11-05,baseline-2025H2-Jan2026,ContextPilot: Fast Long-Context Inference via Context Reuse,https://arxiv.org/abs/2511.03475,False,core,access,model,retrieval_access;representation_organization;evaluation_diagnosis;multi_agent_shared;systems_efficiency,,5,agent-memory
 2511.06179,2025-11-09,baseline-2025H2-Jan2026,MemoriesDB: A Temporal-Semantic-Relational Database for Long-Term Agent Memory / Modeling Experience as a Graph of Temporal-Semantic Surfaces,https://arxiv.org/abs/2511.06179,True,core,experience_learning,model,retrieval_access;representation_organization;experience_skill_learning;systems_efficiency,retrieval_representation;experience_skill_policy,0,agent-memory
 2511.07587,2025-11-10,baseline-2025H2-Jan2026,Beyond Fact Retrieval: Episodic Memory for RAG with Generative Semantic Workspaces,https://arxiv.org/abs/2511.07587,True,core,representation,model,retrieval_access;representation_organization;evaluation_diagnosis;systems_efficiency,retrieval_representation,6,episodic-memory
 2511.08301,2025-11-11,baseline-2025H2-Jan2026,Smarter Together: Creating Agentic Communities of Practice through Shared Experiential Learning,https://arxiv.org/abs/2511.08301,False,core,experience_learning,model,experience_skill_learning;evaluation_diagnosis;multi_agent_shared,experience_skill_policy,5,agent-memory
 2511.12027,2025-11-15,baseline-2025H2-Jan2026,GCAgent: Long-Video Understanding via Schematic and Narrative Episodic Memory,https://arxiv.org/abs/2511.12027,True,core,execution_state,model,retrieval_access;representation_organization;experience_skill_learning;evaluation_diagnosis;multimodal_embodied_gui;systems_efficiency,execution_multimodal,6,episodic-memory
 2511.12997,2025-11-17,baseline-2025H2-Jan2026,WebCoach: Self-Evolving Web Agents with Cross-Session Memory Guidance,https://arxiv.org/abs/2511.12997,True,core,update_forgetting,model,retrieval_access;execution_state;experience_skill_learning;evaluation_diagnosis;multimodal_embodied_gui,experience_skill_policy;execution_multimodal,4,episodic-memory
 2511.16108,2025-11-20,baseline-2025H2-Jan2026,SkyRL-Agent: Efficient RL Training for Multi-turn LLM Agent,https://arxiv.org/abs/2511.16108,False,core,update_forgetting,model,retrieval_access;execution_state;experience_skill_learning;evaluation_diagnosis;multimodal_embodied_gui;systems_efficiency,experience_skill_policy;systems_efficiency,6,agent-memory
 2511.17775,2025-11-21,baseline-2025H2-Jan2026,Episodic Memory in Agentic Frameworks: Suggesting Next Tasks,https://arxiv.org/abs/2511.17775,True,core,update_forgetting,model,retrieval_access;multimodal_embodied_gui,,0,episodic-memory
 2511.21730,2025-11-21,baseline-2025H2-Jan2026,A Benchmark for Procedural Memory Retrieval in Language Agents,https://arxiv.org/abs/2511.21730,True,core,access,model,retrieval_access;representation_organization;experience_skill_learning;evaluation_diagnosis,benchmark_diagnosis;retrieval_representation;experience_skill_policy,5,procedural-memory
 2511.18112,2025-11-22,baseline-2025H2-Jan2026,EchoVLA: Synergistic Declarative Memory for VLA-Driven Mobile Manipulation,https://arxiv.org/abs/2511.18112,True,core,update_forgetting,model,retrieval_access;representation_organization;update_forgetting;execution_state;experience_skill_learning;evaluation_diagnosis;multimodal_embodied_gui,,4,episodic-memory
 2511.18423,2025-11-23,baseline-2025H2-Jan2026,General Agentic Memory Via Deep Research,https://arxiv.org/abs/2511.18423,True,core,update_forgetting,model,retrieval_access;experience_skill_learning;multimodal_embodied_gui,retrieval_representation,3,agent-memory
 2511.19192,2025-11-24,baseline-2025H2-Jan2026,AME: An Efficient Heterogeneous Agentic Memory Engine for Smartphones,https://arxiv.org/abs/2511.19192,True,core,update_forgetting,model,retrieval_access;representation_organization;evaluation_diagnosis;security_privacy_trust;multimodal_embodied_gui;systems_efficiency,systems_efficiency,4,agent-memory
 2511.20297,2025-11-25,baseline-2025H2-Jan2026,Improving Language Agents through BREW,https://arxiv.org/abs/2511.20297,False,core,update_forgetting,model,retrieval_access;representation_organization;execution_state;evaluation_diagnosis;systems_efficiency,,2,agent-memory
 2511.20857,2025-11-25,baseline-2025H2-Jan2026,Evo-Memory: Benchmarking LLM Agent Test-time Learning with Self-Evolving Memory,https://arxiv.org/abs/2511.20857,True,core,update_forgetting,model,retrieval_access;representation_organization;update_forgetting;execution_state;experience_skill_learning;evaluation_diagnosis;multimodal_embodied_gui,benchmark_diagnosis;experience_skill_policy,5,memory-evaluation
 2512.00742,2025-11-30,baseline-2025H2-Jan2026,On the Regulatory Potential of User Interfaces for AI Agent Governance,https://arxiv.org/abs/2512.00742,False,core,access,model,representation_organization;security_privacy_trust,security_governance,0,agent-memory
 2512.02227,2025-12-01,baseline-2025H2-Jan2026,Orchestration Framework for Financial Agents: From Algorithmic Trading to Agentic Trading,https://arxiv.org/abs/2512.02227,False,core,access,model,representation_organization,,1,agent-memory
 2512.02228,2025-12-01,baseline-2025H2-Jan2026,"STRIDE: A Systematic Framework for Selecting AI Modalities -- Agentic AI, AI Assistants, or LLM Calls",https://arxiv.org/abs/2512.02228,False,core,access,model,representation_organization;experience_skill_learning;evaluation_diagnosis;multimodal_embodied_gui;systems_efficiency,,4,long-term-memory
 2512.02425,2025-12-02,baseline-2025H2-Jan2026,WorldMM: Dynamic Multimodal Memory Agent for Long Video Reasoning,https://arxiv.org/abs/2512.02425,True,core,update_forgetting,model,retrieval_access;representation_organization;update_forgetting;evaluation_diagnosis;multimodal_embodied_gui,execution_multimodal,8,agent-memory;episodic-memory
 2512.02458,2025-12-02,baseline-2025H2-Jan2026,Vision to Geometry: 3D Spatial Memory for Sequential Embodied MLLM Reasoning and Exploration,https://arxiv.org/abs/2512.02458,True,core,update_forgetting,model,retrieval_access;evaluation_diagnosis;multimodal_embodied_gui,execution_multimodal,3,agent-memory
 2512.03627,2025-12-03,baseline-2025H2-Jan2026,MemVerse: Multimodal Memory for Lifelong Learning Agents,https://arxiv.org/abs/2512.03627,True,core,experience_learning,model,retrieval_access;representation_organization;update_forgetting;execution_state;experience_skill_learning;evaluation_diagnosis;multimodal_embodied_gui;systems_efficiency,experience_skill_policy;execution_multimodal,4,agent-memory
 2512.04668,2025-12-04,baseline-2025H2-Jan2026,Topology Matters: Measuring Memory Leakage in Multi-Agent LLMs,https://arxiv.org/abs/2512.04668,True,core,execution_state,model,representation_organization;evaluation_diagnosis;security_privacy_trust;multi_agent_shared;multimodal_embodied_gui,security_governance;personal_shared,4,agent-memory
 2512.06688,2025-12-07,baseline-2025H2-Jan2026,PersonaMem-v2: Towards Personalized Intelligence via Learning Implicit User Personas and Agentic Memory,https://arxiv.org/abs/2512.06688,True,core,application,model,experience_skill_learning;systems_efficiency,experience_skill_policy;personal_shared,6,agent-memory
 2512.09458,2025-12-10,baseline-2025H2-Jan2026,Architectures for Building Agentic AI,https://arxiv.org/abs/2512.09458,False,core,execution_state,model,execution_state;security_privacy_trust;multi_agent_shared;multimodal_embodied_gui;systems_efficiency,,1,agent-memory
 2512.10166,2025-12-10,baseline-2025H2-Jan2026,Emergent Collective Memory in Decentralized Multi-Agent AI Systems,https://arxiv.org/abs/2512.10166,True,core,update_forgetting,model,representation_organization;evaluation_diagnosis;multi_agent_shared,personal_shared,7,agent-memory
 2512.10696,2025-12-11,baseline-2025H2-Jan2026,"Remember Me, Refine Me: A Dynamic Procedural Memory Framework for Experience-Driven Agent Evolution",https://arxiv.org/abs/2512.10696,True,core,update_forgetting,model,retrieval_access;representation_organization;experience_skill_learning;systems_efficiency,experience_skill_policy,7,agent-memory;procedural-memory
 2512.11303,2025-12-12,baseline-2025H2-Jan2026,Unifying Dynamic Tool Creation and Cross-Task Experience Sharing through Cognitive Memory Architecture,https://arxiv.org/abs/2512.11303,True,core,update_forgetting,model,retrieval_access;representation_organization;experience_skill_learning;evaluation_diagnosis;multi_agent_shared;systems_efficiency,experience_skill_policy,8,agent-memory;episodic-memory
 2512.12686,2025-12-14,baseline-2025H2-Jan2026,Memoria: A Scalable Agentic Memory Framework for Personalized Conversational AI,https://arxiv.org/abs/2512.12686,True,core,access,model,representation_organization;experience_skill_learning;systems_efficiency,systems_efficiency;personal_shared,0,agent-memory
 2512.12818,2025-12-14,baseline-2025H2-Jan2026,"Hindsight is 20/20: Building Agent Memory that Retains, Recalls, and Reflects",https://arxiv.org/abs/2512.12818,True,core,execution_state,model,retrieval_access;representation_organization;update_forgetting;execution_state;experience_skill_learning;evaluation_diagnosis;security_privacy_trust;multimodal_embodied_gui,retrieval_representation,7,agent-memory;memory-evaluation
 2512.12856,2025-12-14,baseline-2025H2-Jan2026,Forgetful but Faithful: A Cognitive Memory Architecture and Benchmark for Privacy-Aware Generative Agents,https://arxiv.org/abs/2512.12856,True,core,update_forgetting,model,retrieval_access;update_forgetting;evaluation_diagnosis;security_privacy_trust;multimodal_embodied_gui;systems_efficiency,benchmark_diagnosis;update_forgetting;security_governance,2,agent-memory
 2512.12967,2025-12-15,baseline-2025H2-Jan2026,QwenLong-L1.5: Post-Training Recipe for Long-Context Reasoning and Memory Management,https://arxiv.org/abs/2512.12967,True,core,update_forgetting,model,retrieval_access;experience_skill_learning;evaluation_diagnosis;systems_efficiency,,7,agent-memory
 2512.13564,2025-12-15,baseline-2025H2-Jan2026,Memory in the Age of AI Agents,https://arxiv.org/abs/2512.13564,True,core,access,model,retrieval_access;update_forgetting;experience_skill_learning;evaluation_diagnosis;security_privacy_trust;multi_agent_shared;multimodal_embodied_gui;systems_efficiency,,3,agent-memory;memory-evaluation
 2512.16301,2025-12-18,baseline-2025H2-Jan2026,"Adaptation of Agentic AI: A Survey of Post-Training, Memory, and Skills",https://arxiv.org/abs/2512.16301,True,core,experience_learning,model,retrieval_access;experience_skill_learning;evaluation_diagnosis;systems_efficiency,benchmark_diagnosis;experience_skill_policy,4,long-term-memory
 2512.16962,2025-12-18,baseline-2025H2-Jan2026,MemoryGraft: Persistent Compromise of LLM Agents via Poisoned Experience Retrieval,https://arxiv.org/abs/2512.16962,True,core,update_forgetting,model,retrieval_access;experience_skill_learning;evaluation_diagnosis;security_privacy_trust,retrieval_representation;experience_skill_policy;security_governance,3,agent-memory;long-term-memory
 2512.18337,2025-12-20,baseline-2025H2-Jan2026,Towards Efficient Agents: A Co-Design of Inference Architecture and System,https://arxiv.org/abs/2512.18337,False,core,execution_state,model,representation_organization;execution_state;evaluation_diagnosis;systems_efficiency,systems_efficiency,4,agent-memory
 2512.18571,2025-12-21,baseline-2025H2-Jan2026,ESearch-R1: Learning Cost-Aware MLLM Agents for Interactive Embodied Search via Reinforcement Learning,https://arxiv.org/abs/2512.18571,False,core,update_forgetting,model,retrieval_access;execution_state;experience_skill_learning;multimodal_embodied_gui;systems_efficiency,retrieval_representation;experience_skill_policy;execution_multimodal;systems_efficiency,6,episodic-memory
 2512.18746,2025-12-21,baseline-2025H2-Jan2026,MemEvolve: Meta-Evolution of Agent Memory Systems,https://arxiv.org/abs/2512.18746,True,core,update_forgetting,model,retrieval_access;experience_skill_learning;evaluation_diagnosis,experience_skill_policy,2,agent-memory
 2512.18950,2025-12-22,baseline-2025H2-Jan2026,Learning Hierarchical Procedural Memory for LLM Agents through Bayesian Selection and Contrastive Refinement,https://arxiv.org/abs/2512.18950,True,core,update_forgetting,model,representation_organization;update_forgetting;experience_skill_learning;evaluation_diagnosis;systems_efficiency,retrieval_representation;experience_skill_policy,5,procedural-memory
 2512.19537,2025-12-22,baseline-2025H2-Jan2026,Event Extraction in Large Language Model,https://arxiv.org/abs/2512.19537,False,core,execution_state,model,retrieval_access;representation_organization;evaluation_diagnosis;multimodal_embodied_gui,,3,agent-memory
 2512.20111,2025-12-23,baseline-2025H2-Jan2026,ABBEL: Learning Natural-Language Belief States for Memory-Efficient Interaction,https://arxiv.org/abs/2512.20111,False,core,update_forgetting,model,update_forgetting;execution_state;experience_skill_learning;systems_efficiency,experience_skill_policy;systems_efficiency,7,agent-memory
 2512.21567,2025-12-25,baseline-2025H2-Jan2026,Beyond Heuristics: A Decision-Theoretic Framework for Agent Memory Management,https://arxiv.org/abs/2512.21567,True,core,update_forgetting,model,retrieval_access;representation_organization;evaluation_diagnosis;systems_efficiency,,3,agent-memory
 2512.22716,2025-12-27,baseline-2025H2-Jan2026,Memento 2: Learning by Stateful Reflective Memory,https://arxiv.org/abs/2512.22716,True,core,update_forgetting,model,retrieval_access;execution_state;experience_skill_learning;evaluation_diagnosis;multimodal_embodied_gui;systems_efficiency,experience_skill_policy,2,episodic-memory
 2512.23343,2025-12-29,baseline-2025H2-Jan2026,AI Meets Brain: Memory Systems from Cognitive Neuroscience to Autonomous Agents,https://arxiv.org/abs/2512.23343,True,core,representation,model,retrieval_access;execution_state;experience_skill_learning;evaluation_diagnosis;security_privacy_trust;multimodal_embodied_gui;systems_efficiency,,2,agent-memory
 2601.01885,2026-01-05,baseline-2025H2-Jan2026,Agentic Memory: Learning Unified Long-Term and Short-Term Memory Management for Large Language Model Agents,https://arxiv.org/abs/2601.01885,True,core,update_forgetting,model,retrieval_access;update_forgetting;execution_state;experience_skill_learning;evaluation_diagnosis;systems_efficiency,experience_skill_policy,5,agent-memory;long-term-memory
 2601.02577,2026-01-05,baseline-2025H2-Jan2026,Orchestral AI: A Framework for Agent Orchestration,https://arxiv.org/abs/2601.02577,False,core,representation,model,representation_organization;systems_efficiency,,0,agent-memory
 2601.02732,2026-01-06,baseline-2025H2-Jan2026,Agentic Memory Enhanced Recursive Reasoning for Root Cause Localization in Microservices,https://arxiv.org/abs/2601.02732,True,core,update_forgetting,model,multi_agent_shared;systems_efficiency,,5,agent-memory
 2601.02744,2026-01-06,baseline-2025H2-Jan2026,SYNAPSE: Empowering LLM Agents with Episodic-Semantic Memory via Spreading Activation,https://arxiv.org/abs/2601.02744,True,core,update_forgetting,model,retrieval_access;evaluation_diagnosis,,6,agent-memory
 2601.03236,2026-01-06,baseline-2025H2-Jan2026,MAGMA: A Multi-Graph based Agentic Memory Architecture for AI Agents,https://arxiv.org/abs/2601.03236,True,core,update_forgetting,model,retrieval_access;representation_organization;execution_state;multimodal_embodied_gui,retrieval_representation,4,agent-memory
 2601.03543,2026-01-07,baseline-2025H2-Jan2026,EvolMem: A Cognitive-Driven Benchmark for Multi-Session Dialogue Memory,https://arxiv.org/abs/2601.03543,True,core,representation,model,evaluation_diagnosis;multimodal_embodied_gui,benchmark_diagnosis,6,agent-memory
 2601.03785,2026-01-07,baseline-2025H2-Jan2026,Membox: Weaving Topic Continuity into Long-Range Memory for LLM Agents,https://arxiv.org/abs/2601.03785,True,core,experience_learning,model,retrieval_access;representation_organization;update_forgetting,,4,agent-memory
 2601.04170,2026-01-07,baseline-2025H2-Jan2026,Agent Drift: Quantifying Behavioral Degradation in Multi-Agent LLM Systems Over Extended Interactions,https://arxiv.org/abs/2601.04170,False,supporting,execution_state,model,retrieval_access;representation_organization;update_forgetting;evaluation_diagnosis;multi_agent_shared;systems_efficiency,personal_shared,1,episodic-memory
 2601.05504,2026-01-09,baseline-2025H2-Jan2026,Memory Poisoning Attack and Defense on Memory Based LLM-Agents,https://arxiv.org/abs/2601.05504,True,supporting,security_privacy,model,retrieval_access;evaluation_diagnosis;security_privacy_trust,security_governance,6,long-term-memory
 2601.06282,2026-01-09,baseline-2025H2-Jan2026,Amory: Building Coherent Narrative-Driven Agent Memory through Agentic Reasoning,https://arxiv.org/abs/2601.06282,True,core,experience_learning,model,retrieval_access;representation_organization;update_forgetting;evaluation_diagnosis;systems_efficiency,,8,agent-memory
 2601.06377,2026-01-10,baseline-2025H2-Jan2026,HiMem: Hierarchical Long-Term Memory for LLM Long-Horizon Agents,https://arxiv.org/abs/2601.06377,True,core,experience_learning,model,retrieval_access;representation_organization;update_forgetting;execution_state;experience_skill_learning;evaluation_diagnosis;systems_efficiency,retrieval_representation;execution_multimodal,6,episodic-memory
 2601.06411,2026-01-10,baseline-2025H2-Jan2026,Structured Episodic Event Memory,https://arxiv.org/abs/2601.06411,True,core,experience_learning,model,retrieval_access;representation_organization;evaluation_diagnosis;security_privacy_trust,retrieval_representation,6,episodic-memory
 2601.10744,2026-01-11,baseline-2025H2-Jan2026,Explore with Long-term Memory: A Benchmark and Multimodal LLM-based Reinforcement Learning Framework for Embodied Exploration,https://arxiv.org/abs/2601.10744,True,supporting,execution_state,model,retrieval_access;execution_state;experience_skill_learning;evaluation_diagnosis;multimodal_embodied_gui,benchmark_diagnosis;experience_skill_policy;execution_multimodal,4,agent-memory;episodic-memory
 2601.07470,2026-01-12,baseline-2025H2-Jan2026,Learning How to Remember: A Meta-Cognitive Management Method for Structured and Transferable Agent Memory,https://arxiv.org/abs/2601.07470,True,core,experience_learning,model,representation_organization;execution_state;experience_skill_learning,retrieval_representation;experience_skill_policy,4,agent-memory
 2601.07779,2026-01-12,baseline-2025H2-Jan2026,OS-Symphony: A Holistic Framework for Robust and Generalist Computer-Using Agent,https://arxiv.org/abs/2601.07779,False,supporting,execution_state,model,retrieval_access;execution_state;experience_skill_learning;evaluation_diagnosis;multimodal_embodied_gui,,6,agent-memory
 2601.08160,2026-01-13,baseline-2025H2-Jan2026,SwiftMem: Fast Agentic Memory via Query-aware Indexing,https://arxiv.org/abs/2601.08160,True,core,experience_learning,model,retrieval_access;representation_organization;update_forgetting;evaluation_diagnosis;systems_efficiency,retrieval_representation,5,agent-memory
 2601.08323,2026-01-13,baseline-2025H2-Jan2026,AtomMem : Learnable Dynamic Agentic Memory with Atomic Memory Operation,https://arxiv.org/abs/2601.08323,True,core,experience_learning,model,representation_organization;update_forgetting;execution_state;experience_skill_learning;evaluation_diagnosis,experience_skill_policy,4,agent-memory
 2601.09113,2026-01-14,baseline-2025H2-Jan2026,The AI Hippocampus: How Far are We From Human Memory?,https://arxiv.org/abs/2601.09113,True,supporting,experience_learning,model,retrieval_access;representation_organization;execution_state;experience_skill_learning;evaluation_diagnosis;multi_agent_shared;multimodal_embodied_gui;systems_efficiency,benchmark_diagnosis,2,agent-memory
 2601.09636,2026-01-14,baseline-2025H2-Jan2026,PersonalAlign: Hierarchical Implicit Intent Alignment for Personalized GUI Agent with Long-Term User-Centric Records,https://arxiv.org/abs/2601.09636,False,core,execution_state,model,representation_organization;evaluation_diagnosis;multimodal_embodied_gui,retrieval_representation;execution_multimodal;personal_shared,5,agent-memory
 2602.06051,2026-01-14,baseline-2025H2-Jan2026,CAST: Character-and-Scene Episodic Memory for Agents,https://arxiv.org/abs/2602.06051,True,core,representation,model,retrieval_access;representation_organization;experience_skill_learning,,6,agent-memory;episodic-memory
 2602.06052,2026-01-14,baseline-2025H2-Jan2026,Rethinking Memory Mechanisms of Foundation Agents in the Second Half: A Survey,https://arxiv.org/abs/2602.06052,True,core,access,model,retrieval_access;execution_state;experience_skill_learning;evaluation_diagnosis,benchmark_diagnosis,2,agent-memory
 2601.10702,2026-01-15,baseline-2025H2-Jan2026,Grounding Agent Memory in Contextual Intent,https://arxiv.org/abs/2601.10702,True,core,update_forgetting,model,retrieval_access;representation_organization;execution_state;evaluation_diagnosis,,6,agent-memory
 2601.11854,2026-01-17,baseline-2025H2-Jan2026,ATOD: An Evaluation Framework and Benchmark for Agentic Task-Oriented Dialogue Systems,https://arxiv.org/abs/2601.11854,False,core,representation,model,execution_state;evaluation_diagnosis,benchmark_diagnosis,4,agent-memory
 2601.12771,2026-01-19,baseline-2025H2-Jan2026,Who Does This Name Remind You of ? Nationality Prediction via Large Language Model Associative Memory,https://arxiv.org/abs/2601.12771,True,core,update_forgetting,model,retrieval_access;multi_agent_shared;multimodal_embodied_gui,,5,agent-memory
 2601.14192,2026-01-20,baseline-2025H2-Jan2026,"Toward Efficient Agents: Memory, Tool learning, and Planning",https://arxiv.org/abs/2601.14192,True,core,update_forgetting,model,retrieval_access;update_forgetting;execution_state;experience_skill_learning;evaluation_diagnosis;systems_efficiency,experience_skill_policy;systems_efficiency,3,agent-memory
 2601.14735,2026-01-21,baseline-2025H2-Jan2026,Optimizing FaaS Platforms for MCP-enabled Agentic Workflows,https://arxiv.org/abs/2601.14735,False,core,execution_state,model,retrieval_access;evaluation_diagnosis;multi_agent_shared;systems_efficiency,,4,agent-memory
 2601.16690,2026-01-23,baseline-2025H2-Jan2026,EMemBench: Interactive Benchmarking of Episodic Memory for VLM Agents,https://arxiv.org/abs/2601.16690,True,core,experience_learning,model,retrieval_access;execution_state;experience_skill_learning;evaluation_diagnosis;security_privacy_trust;multimodal_embodied_gui,benchmark_diagnosis,6,agent-memory
 2601.16872,2026-01-23,baseline-2025H2-Jan2026,From Atom to Community: Structured and Evolving Agent Memory for User Behavior Modeling,https://arxiv.org/abs/2601.16872,True,core,update_forgetting,model,representation_organization;update_forgetting,retrieval_representation;experience_skill_policy,4,agent-memory
 2601.17887,2026-01-25,baseline-2025H2-Jan2026,When Personalization Legitimizes Risks: Uncovering Safety Vulnerabilities in Personalized Dialogue Agents,https://arxiv.org/abs/2601.17887,False,core,execution_state,model,representation_organization;experience_skill_learning;evaluation_diagnosis;security_privacy_trust,security_governance;personal_shared,6,long-term-memory
 2601.18642,2026-01-26,baseline-2025H2-Jan2026,FadeMem: Biologically-Inspired Forgetting for Efficient Agent Memory,https://arxiv.org/abs/2601.18642,True,core,update_forgetting,model,retrieval_access;representation_organization;update_forgetting;security_privacy_trust;multimodal_embodied_gui;systems_efficiency,update_forgetting;systems_efficiency,3,agent-memory
 2601.20352,2026-01-28,baseline-2025H2-Jan2026,AMA: Adaptive Memory via Multi-Agent Collaboration,https://arxiv.org/abs/2601.20352,True,core,update_forgetting,model,retrieval_access;representation_organization;update_forgetting;evaluation_diagnosis;multi_agent_shared;systems_efficiency,experience_skill_policy;personal_shared,7,agent-memory;long-term-memory
 2601.20465,2026-01-28,baseline-2025H2-Jan2026,BMAM: Brain-inspired Multi-Agent Memory Framework,https://arxiv.org/abs/2601.20465,True,core,execution_state,model,retrieval_access;representation_organization;execution_state;evaluation_diagnosis;multi_agent_shared;systems_efficiency,personal_shared,4,agent-memory
 2601.21468,2026-01-29,baseline-2025H2-Jan2026,MemOCR: Layout-Aware Visual Memory for Efficient Long-Horizon Reasoning,https://arxiv.org/abs/2601.21468,True,core,update_forgetting,model,representation_organization;execution_state;experience_skill_learning;evaluation_diagnosis;multimodal_embodied_gui;systems_efficiency,execution_multimodal;systems_efficiency,4,agent-memory
 2601.21714,2026-01-29,baseline-2025H2-Jan2026,E-mem: Multi-agent based Episodic Context Reconstruction for LLM Agent Memory,https://arxiv.org/abs/2601.21714,True,core,update_forgetting,model,retrieval_access;representation_organization;execution_state;evaluation_diagnosis;security_privacy_trust;multi_agent_shared;systems_efficiency,personal_shared,2,agent-memory
 2601.21841,2026-01-29,baseline-2025H2-Jan2026,Embodied Task Planning via Graph-Informed Action Generation with Large Language Models,https://arxiv.org/abs/2601.21841,False,core,update_forgetting,model,retrieval_access;representation_organization;execution_state;experience_skill_learning;evaluation_diagnosis;multimodal_embodied_gui;systems_efficiency,retrieval_representation;execution_multimodal,7,agent-memory
 2601.22964,2026-01-30,baseline-2025H2-Jan2026,EvoClinician: A Self-Evolving Agent for Multi-Turn Medical Diagnosis via Test-Time Evolutionary Learning,https://arxiv.org/abs/2601.22964,False,core,update_forgetting,model,update_forgetting;experience_skill_learning;evaluation_diagnosis;systems_efficiency,benchmark_diagnosis;experience_skill_policy,5,agent-memory
 2601.22974,2026-01-30,baseline-2025H2-Jan2026,MiTa: A Hierarchical Multi-Agent Collaboration Framework with Memory-integrated and Task Allocation,https://arxiv.org/abs/2601.22974,True,core,update_forgetting,model,representation_organization;update_forgetting;execution_state;multi_agent_shared;multimodal_embodied_gui,retrieval_representation;personal_shared,4,episodic-memory
 2601.23014,2026-01-30,baseline-2025H2-Jan2026,Mem-T: Densifying Rewards for Long-Horizon Memory Agents,https://arxiv.org/abs/2601.23014,True,core,update_forgetting,model,retrieval_access;representation_organization;update_forgetting;execution_state;experience_skill_learning;multimodal_embodied_gui;systems_efficiency,execution_multimodal,3,agent-memory
 2602.00352,2026-01-30,baseline-2025H2-Jan2026,DETOUR: An Interactive Benchmark for Dual-Agent Search and Reasoning,https://arxiv.org/abs/2602.00352,False,core,access,model,retrieval_access;evaluation_diagnosis;multimodal_embodied_gui,benchmark_diagnosis;retrieval_representation,4,agent-memory
 2602.00364,2026-01-30,baseline-2025H2-Jan2026,"""Someone Hid It"": Query-Agnostic Black-Box Attacks on LLM-Based Retrieval",https://arxiv.org/abs/2602.00364,False,core,access,model,retrieval_access;evaluation_diagnosis;security_privacy_trust;systems_efficiency,retrieval_representation;security_governance,2,agent-memory
 2602.00675,2026-01-31,baseline-2025H2-Jan2026,Factored Reasoning with Inner Speech and Persistent Memory for Evidence-Grounded Human-Robot Interaction,https://arxiv.org/abs/2602.00675,True,core,execution_state,model,retrieval_access;representation_organization;update_forgetting;evaluation_diagnosis;multimodal_embodied_gui;systems_efficiency,execution_multimodal,3,agent-memory
 2602.01869,2026-02-02,transition-2026-Feb-Mar,Skill-Pro: Learning Reusable Skills from Experience via Non-Parametric PPO for LLM Agents,https://arxiv.org/abs/2602.01869,False,core,update_forgetting,model,update_forgetting;experience_skill_learning;evaluation_diagnosis;multimodal_embodied_gui;systems_efficiency,experience_skill_policy,3,procedural-memory
 2602.02007,2026-02-02,transition-2026-Feb-Mar,Beyond RAG for Agent Memory: Retrieval by Decoupling and Aggregation,https://arxiv.org/abs/2602.02007,True,core,update_forgetting,model,retrieval_access;representation_organization;update_forgetting;multimodal_embodied_gui;systems_efficiency,retrieval_representation,2,agent-memory
 2602.02164,2026-02-02,transition-2026-Feb-Mar,Co-RedTeam: Orchestrated Security Discovery and Exploitation with LLM Agents,https://arxiv.org/abs/2602.02164,False,supporting,security_privacy,model,representation_organization;experience_skill_learning;evaluation_diagnosis;security_privacy_trust;multi_agent_shared,security_governance,7,long-term-memory
 2602.02369,2026-02-02,transition-2026-Feb-Mar,Live-Evo: Online Evolution of Agentic Memory from Continuous Feedback,https://arxiv.org/abs/2602.02369,True,core,update_forgetting,model,retrieval_access;update_forgetting;experience_skill_learning;evaluation_diagnosis;multimodal_embodied_gui,,5,agent-memory
 2602.02474,2026-02-02,transition-2026-Feb-Mar,MemSkill: Learning and Evolving Memory Skills for Self-Evolving Agents,https://arxiv.org/abs/2602.02474,True,core,update_forgetting,model,representation_organization;update_forgetting;experience_skill_learning;multimodal_embodied_gui;systems_efficiency,experience_skill_policy,5,agent-memory
 2602.03036,2026-02-03,transition-2026-Feb-Mar,LatentMem: Customizing Latent Memory for Multi-Agent Systems,https://arxiv.org/abs/2602.03036,True,core,update_forgetting,model,retrieval_access;representation_organization;experience_skill_learning;evaluation_diagnosis;multi_agent_shared;systems_efficiency,personal_shared,5,agent-memory
 2602.03224,2026-02-03,transition-2026-Feb-Mar,TAME: A Trustworthy Test-Time Evolution of Agent Memory with Systematic Benchmarking,https://arxiv.org/abs/2602.03224,True,core,security_privacy,model,retrieval_access;update_forgetting;experience_skill_learning;evaluation_diagnosis;security_privacy_trust;multi_agent_shared;multimodal_embodied_gui,benchmark_diagnosis;security_governance,2,agent-memory
 2602.03315,2026-02-03,transition-2026-Feb-Mar,Memora: A Harmonic Memory Representation Balancing Abstraction and Specificity,https://arxiv.org/abs/2602.03315,True,core,update_forgetting,model,retrieval_access;representation_organization;update_forgetting;evaluation_diagnosis;systems_efficiency,retrieval_representation,2,agent-memory
 2602.04482,2026-02-04,transition-2026-Feb-Mar,ProAgentBench: Evaluating LLM Agents for Proactive Assistance with Real-World Data,https://arxiv.org/abs/2602.04482,False,supporting,application,model,representation_organization;evaluation_diagnosis;security_privacy_trust;systems_efficiency,benchmark_diagnosis,7,long-term-memory
 2602.04640,2026-02-04,transition-2026-Feb-Mar,"Towards Structured, State-Aware, and Execution-Grounded Reasoning for Software Engineering Agents",https://arxiv.org/abs/2602.04640,False,core,update_forgetting,model,representation_organization;execution_state,retrieval_representation,0,agent-memory
 2603.00026,2026-02-04,transition-2026-Feb-Mar,ActMem: Bridging the Gap Between Memory Retrieval and Reasoning in LLM Agents,https://arxiv.org/abs/2603.00026,True,core,update_forgetting,model,retrieval_access;representation_organization;update_forgetting;evaluation_diagnosis,retrieval_representation,6,memory-evaluation
 2602.05665,2026-02-05,transition-2026-Feb-Mar,"Graph-based Agent Memory: Taxonomy, Techniques, and Applications",https://arxiv.org/abs/2602.05665,True,core,update_forgetting,model,retrieval_access;representation_organization;execution_state;experience_skill_learning;evaluation_diagnosis;systems_efficiency,benchmark_diagnosis;retrieval_representation,5,agent-memory
 2602.06025,2026-02-05,transition-2026-Feb-Mar,Learning Query-Aware Budget-Tier Routing for Runtime Agent Memory,https://arxiv.org/abs/2602.06025,True,supporting,application,model,retrieval_access;representation_organization;experience_skill_learning;systems_efficiency,experience_skill_policy;systems_efficiency,2,agent-memory
 2602.17692,2026-02-06,transition-2026-Feb-Mar,Agentic Unlearning: When LLM Agent Meets Machine Unlearning,https://arxiv.org/abs/2602.17692,False,core,update_forgetting,model,retrieval_access;update_forgetting;evaluation_diagnosis;multimodal_embodied_gui,experience_skill_policy;security_governance,2,long-term-memory
 2603.03296,2026-02-06,transition-2026-Feb-Mar,PlugMem: A Task-Agnostic Plugin Memory Module for LLM Agents,https://arxiv.org/abs/2603.03296,True,core,update_forgetting,model,retrieval_access;representation_organization;execution_state;experience_skill_learning;evaluation_diagnosis;multimodal_embodied_gui;systems_efficiency,,5,episodic-memory;long-term-memory
 2602.07398,2026-02-07,transition-2026-Feb-Mar,AgentSys: Secure and Dynamic LLM Agents Through Explicit Hierarchical Memory Management,https://arxiv.org/abs/2602.07398,True,core,update_forgetting,model,representation_organization;evaluation_diagnosis;security_privacy_trust,retrieval_representation,6,agent-memory
 2602.07624,2026-02-07,transition-2026-Feb-Mar,M2A: Multimodal Memory Agent with Dual-Layer Hybrid Memory for Long-Term Personalized Interactions,https://arxiv.org/abs/2602.07624,True,core,update_forgetting,model,retrieval_access;update_forgetting;multimodal_embodied_gui;systems_efficiency,execution_multimodal;personal_shared,6,agent-memory
 2602.07755,2026-02-08,transition-2026-Feb-Mar,Learning to Continually Learn via Meta-learning Agentic Memory Designs,https://arxiv.org/abs/2602.07755,True,core,update_forgetting,model,retrieval_access;representation_organization;update_forgetting;execution_state;experience_skill_learning;evaluation_diagnosis;systems_efficiency,experience_skill_policy,3,agent-memory
 2602.07885,2026-02-08,transition-2026-Feb-Mar,MemFly: On-the-Fly Memory Optimization via Information Bottleneck,https://arxiv.org/abs/2602.07885,True,core,update_forgetting,model,retrieval_access;representation_organization;systems_efficiency,,4,long-term-memory
 2602.08369,2026-02-09,transition-2026-Feb-Mar,MemAdapter: Fast Alignment across Agent Memory Paradigms via Generative Subgraph Retrieval,https://arxiv.org/abs/2602.08369,True,core,update_forgetting,model,retrieval_access;execution_state;evaluation_diagnosis;systems_efficiency,retrieval_representation;experience_skill_policy,6,agent-memory
 2602.09319,2026-02-10,transition-2026-Feb-Mar,Benchmarking Knowledge-Extraction Attack and Defense on Retrieval-Augmented Generation,https://arxiv.org/abs/2602.09319,False,core,access,model,retrieval_access;representation_organization;update_forgetting;evaluation_diagnosis;security_privacy_trust;systems_efficiency,benchmark_diagnosis;retrieval_representation;security_governance;systems_efficiency,2,agent-memory
 2602.09712,2026-02-10,transition-2026-Feb-Mar,TraceMem: Weaving Narrative Memory Schemata from User Conversational Traces,https://arxiv.org/abs/2602.09712,True,core,execution_state,model,retrieval_access;representation_organization;update_forgetting;evaluation_diagnosis,,6,episodic-memory
 2602.10715,2026-02-11,transition-2026-Feb-Mar,Locomo-Plus: Beyond-Factual Cognitive Memory Evaluation Framework for LLM Agents,https://arxiv.org/abs/2602.10715,True,core,execution_state,model,retrieval_access;evaluation_diagnosis,benchmark_diagnosis,4,memory-evaluation
 2602.11243,2026-02-11,transition-2026-Feb-Mar,Evaluating Memory Structure in LLM Agents,https://arxiv.org/abs/2602.11243,True,core,representation,model,retrieval_access;representation_organization;update_forgetting;evaluation_diagnosis;multimodal_embodied_gui,benchmark_diagnosis;retrieval_representation,3,agent-memory;long-term-memory;memory-evaluation
 2602.13530,2026-02-13,transition-2026-Feb-Mar,REMem: Reasoning with Episodic Memory in Language Agent,https://arxiv.org/abs/2602.13530,True,core,representation,model,retrieval_access;representation_organization;experience_skill_learning;evaluation_diagnosis,,6,memory-evaluation
 2602.18493,2026-02-13,transition-2026-Feb-Mar,Learning to Remember: End-to-End Training of Memory Agents for Long-Context Reasoning,https://arxiv.org/abs/2602.18493,True,core,execution_state,model,retrieval_access;representation_organization;update_forgetting;execution_state;experience_skill_learning;evaluation_diagnosis,experience_skill_policy,4,agent-memory
 2602.13594,2026-02-14,transition-2026-Feb-Mar,Hippocampus: An Efficient and Scalable Memory Module for Agentic AI,https://arxiv.org/abs/2602.13594,True,core,execution_state,model,retrieval_access;representation_organization;execution_state;evaluation_diagnosis;systems_efficiency,systems_efficiency,2,agent-memory
 2602.13933,2026-02-15,transition-2026-Feb-Mar,HyMem: Hybrid Memory Architecture with Dynamic Retrieval Scheduling,https://arxiv.org/abs/2602.13933,True,core,execution_state,model,retrieval_access;representation_organization;experience_skill_learning;evaluation_diagnosis;systems_efficiency,retrieval_representation,6,long-term-memory
 2602.14038,2026-02-15,transition-2026-Feb-Mar,Choosing How to Remember: Adaptive Memory Structures for LLM Agents,https://arxiv.org/abs/2602.14038,True,core,execution_state,model,representation_organization;execution_state;evaluation_diagnosis,retrieval_representation;experience_skill_policy,4,agent-memory
 2602.15274,2026-02-17,transition-2026-Feb-Mar,When Remembering and Planning are Worth it: Navigating under Change,https://arxiv.org/abs/2602.15274,True,core,execution_state,model,retrieval_access;execution_state;experience_skill_learning;multimodal_embodied_gui;systems_efficiency,,0,agent-memory
 2602.15513,2026-02-17,transition-2026-Feb-Mar,HIMM: Human-Inspired Long-Term Memory Modeling for Embodied Exploration and Question Answering,https://arxiv.org/abs/2602.15513,True,core,execution_state,model,retrieval_access;representation_organization;execution_state;experience_skill_learning;evaluation_diagnosis;multimodal_embodied_gui;systems_efficiency,execution_multimodal,4,episodic-memory
 2602.15654,2026-02-17,transition-2026-Feb-Mar,Zombie Agents: Persistent Control of Self-Evolving LLM Agents via Self-Reinforcing Injections,https://arxiv.org/abs/2602.15654,False,core,execution_state,model,retrieval_access;update_forgetting;execution_state;experience_skill_learning;evaluation_diagnosis;security_privacy_trust;systems_efficiency,experience_skill_policy,3,long-term-memory
 2603.02240,2026-02-17,transition-2026-Feb-Mar,SuperLocalMemory: Privacy-Preserving Multi-Agent Memory with Bayesian Trust Defense Against Memory Poisoning,https://arxiv.org/abs/2603.02240,True,core,execution_state,model,retrieval_access;representation_organization;evaluation_diagnosis;security_privacy_trust;multi_agent_shared;systems_efficiency,security_governance;systems_efficiency;personal_shared,5,agent-memory
 2603.04428,2026-02-17,transition-2026-Feb-Mar,Agent Memory Below the Prompt: Persistent Q4 KV Cache for Multi-Agent LLM Inference on Edge Devices,https://arxiv.org/abs/2603.04428,True,core,execution_state,model,evaluation_diagnosis;multi_agent_shared;systems_efficiency,systems_efficiency;personal_shared,5,agent-memory
 2602.16313,2026-02-18,transition-2026-Feb-Mar,MemoryArena: Benchmarking Agent Memory in Interdependent Multi-Session Agentic Tasks,https://arxiv.org/abs/2602.16313,True,core,execution_state,model,retrieval_access;execution_state;experience_skill_learning;evaluation_diagnosis;multimodal_embodied_gui,benchmark_diagnosis,3,agent-memory;memory-evaluation
 2602.16493,2026-02-18,transition-2026-Feb-Mar,MMA: Multimodal Memory Agent,https://arxiv.org/abs/2602.16493,True,core,execution_state,model,retrieval_access;representation_organization;update_forgetting;execution_state;evaluation_diagnosis;multimodal_embodied_gui,execution_multimodal,8,agent-memory
 2602.19320,2026-02-22,transition-2026-Feb-Mar,Anatomy of Agentic Memory: Taxonomy and Empirical Analysis of Evaluation and System Limitations,https://arxiv.org/abs/2602.19320,True,core,execution_state,model,representation_organization;execution_state;evaluation_diagnosis;systems_efficiency,benchmark_diagnosis,4,agent-memory
 2603.04443,2026-02-22,transition-2026-Feb-Mar,AMV-L: Lifecycle-Managed Agent Memory for Tail-Latency Control in Long-Running LLM Systems,https://arxiv.org/abs/2603.04443,True,core,execution_state,model,retrieval_access;update_forgetting;evaluation_diagnosis;systems_efficiency,update_forgetting;systems_efficiency,7,agent-memory;long-term-memory
 2602.21394,2026-02-24,transition-2026-Feb-Mar,MemoPhishAgent: Memory-Augmented Multi-Modal LLM Agent for Phishing URL Detection,https://arxiv.org/abs/2602.21394,True,core,execution_state,model,retrieval_access;experience_skill_learning;evaluation_diagnosis;security_privacy_trust;multimodal_embodied_gui,,8,episodic-memory
 2602.21477,2026-02-25,transition-2026-Feb-Mar,Pancake: Hierarchical Memory System for Multi-Agent LLM Serving,https://arxiv.org/abs/2602.21477,True,core,update_forgetting,model,retrieval_access;representation_organization;update_forgetting;multi_agent_shared;systems_efficiency,retrieval_representation;systems_efficiency;personal_shared,6,agent-memory
 2602.22406,2026-02-25,transition-2026-Feb-Mar,Towards Autonomous Memory Agents,https://arxiv.org/abs/2602.22406,True,core,update_forgetting,model,retrieval_access;update_forgetting;experience_skill_learning;evaluation_diagnosis;systems_efficiency,,6,agent-memory
 2602.22769,2026-02-26,transition-2026-Feb-Mar,AMA-Bench: Evaluating Long-Horizon Memory for Agentic Applications,https://arxiv.org/abs/2602.22769,True,core,update_forgetting,model,retrieval_access;execution_state;evaluation_diagnosis,benchmark_diagnosis;execution_multimodal,8,agent-memory;memory-evaluation
 2602.23720,2026-02-27,transition-2026-Feb-Mar,The Auton Agentic AI Framework,https://arxiv.org/abs/2602.23720,False,core,execution_state,model,representation_organization;update_forgetting;experience_skill_learning;security_privacy_trust;systems_efficiency,,0,episodic-memory
 2602.23937,2026-02-27,transition-2026-Feb-Mar,Enhancing Vision-Language Navigation with Multimodal Event Knowledge from Real-World Indoor Tour Videos,https://arxiv.org/abs/2602.23937,False,core,execution_state,model,retrieval_access;representation_organization;execution_state;evaluation_diagnosis;multimodal_embodied_gui,execution_multimodal;systems_efficiency,4,episodic-memory
 2603.01160,2026-03-01,transition-2026-Feb-Mar,Semantic XPath: Structured Agentic Memory Access for Conversational AI,https://arxiv.org/abs/2603.01160,True,core,access,model,retrieval_access;representation_organization;update_forgetting;multimodal_embodied_gui;systems_efficiency,retrieval_representation,4,agent-memory
 2603.01966,2026-03-02,transition-2026-Feb-Mar,AMemGym: Interactive Memory Benchmarking for Assistants in Long-Horizon Conversations,https://arxiv.org/abs/2603.01966,True,core,update_forgetting,model,representation_organization;execution_state;evaluation_diagnosis;multimodal_embodied_gui;systems_efficiency,benchmark_diagnosis;execution_multimodal,2,agent-memory;memory-evaluation
 2603.02206,2026-03-02,transition-2026-Feb-Mar,VoiceAgentRAG: Solving the RAG Latency Bottleneck in Real-Time Voice Agents Using Dual-Agent Architectures,https://arxiv.org/abs/2603.02206,False,core,access,model,retrieval_access;representation_organization;systems_efficiency,retrieval_representation;systems_efficiency,1,agent-memory
 2603.02473,2026-03-02,transition-2026-Feb-Mar,Diagnosing Retrieval vs. Utilization Bottlenecks in LLM Agent Memory,https://arxiv.org/abs/2603.02473,True,core,access,model,retrieval_access;evaluation_diagnosis;systems_efficiency,benchmark_diagnosis;retrieval_representation,6,agent-memory
 2603.03024,2026-03-03,transition-2026-Feb-Mar,MA-CoNav: A Master-Slave Multi-Agent Framework with Hierarchical Collaboration and Dual-Level Reflection for Long-Horizon Embodied VLN,https://arxiv.org/abs/2603.03024,False,core,update_forgetting,model,representation_organization;execution_state;experience_skill_learning;evaluation_diagnosis;multi_agent_shared;multimodal_embodied_gui;systems_efficiency,retrieval_representation;execution_multimodal;personal_shared,4,agent-memory
 2603.03148,2026-03-03,transition-2026-Feb-Mar,From Language to Action: Can LLM-Based Agents Be Used for Embodied Robot Cognition?,https://arxiv.org/abs/2603.03148,False,core,execution_state,model,representation_organization;execution_state;experience_skill_learning;evaluation_diagnosis;multimodal_embodied_gui,execution_multimodal,4,episodic-memory
 2603.15642,2026-03-03,transition-2026-Feb-Mar,CraniMem: Cranial Inspired Gated and Bounded Memory for Agentic Systems,https://arxiv.org/abs/2603.15642,True,core,update_forgetting,model,retrieval_access;representation_organization;update_forgetting;execution_state;evaluation_diagnosis,,5,agent-memory
 2603.04549,2026-03-04,transition-2026-Feb-Mar,Adaptive Memory Admission Control for LLM Agents,https://arxiv.org/abs/2603.04549,True,core,update_forgetting,model,retrieval_access;representation_organization;update_forgetting;evaluation_diagnosis;systems_efficiency,experience_skill_policy,7,long-term-memory
 2603.04740,2026-03-05,transition-2026-Feb-Mar,Memory as Ontology: A Constitutional Memory Architecture for Persistent Digital Citizens,https://arxiv.org/abs/2603.04740,True,core,execution_state,model,retrieval_access;representation_organization;security_privacy_trust;multimodal_embodied_gui;systems_efficiency,retrieval_representation,0,agent-memory
 2603.23516,2026-03-06,transition-2026-Feb-Mar,MSA: Memory Sparse Attention for Efficient End-to-End Memory Model Scaling to 100M Tokens,https://arxiv.org/abs/2603.23516,True,core,update_forgetting,model,evaluation_diagnosis;systems_efficiency,systems_efficiency,5,agent-memory
 2603.08755,2026-03-07,transition-2026-Feb-Mar,Turn: A Language for Agentic Computation,https://arxiv.org/abs/2603.08755,False,core,update_forgetting,model,evaluation_diagnosis,,3,agent-memory
 2603.07392,2026-03-08,transition-2026-Feb-Mar,Can Large Language Models Keep Up? Benchmarking Online Adaptation to Continual Knowledge Streams,https://arxiv.org/abs/2603.07392,False,core,update_forgetting,model,representation_organization;evaluation_diagnosis,benchmark_diagnosis;experience_skill_policy;systems_efficiency,4,agent-memory
 2603.07670,2026-03-08,transition-2026-Feb-Mar,"Memory for Autonomous LLM Agents:Mechanisms, Evaluation, and Emerging Frontiers",https://arxiv.org/abs/2603.07670,True,core,representation,model,retrieval_access;representation_organization;update_forgetting;experience_skill_learning;evaluation_diagnosis;security_privacy_trust;multi_agent_shared;multimodal_embodied_gui;systems_efficiency,benchmark_diagnosis,2,agent-memory
 2603.07784,2026-03-08,transition-2026-Feb-Mar,ProgAgent:A Continual RL Agent with Progress-Aware Rewards,https://arxiv.org/abs/2603.07784,False,core,update_forgetting,model,update_forgetting;experience_skill_learning;evaluation_diagnosis;security_privacy_trust;multimodal_embodied_gui;systems_efficiency,experience_skill_policy,4,agent-memory
 2603.10062,2026-03-09,transition-2026-Feb-Mar,Multi-Agent Memory from a Computer Architecture Perspective: Visions and Challenges Ahead,https://arxiv.org/abs/2603.10062,True,core,execution_state,model,representation_organization;multi_agent_shared;multimodal_embodied_gui,personal_shared,0,agent-memory
 2604.00009,2026-03-09,transition-2026-Feb-Mar,"Eyla: Toward an Identity-Anchored LLM Architecture with Integrated Biological Priors -- Vision, Implementation Attempt, and Lessons from AI-Assisted Development",https://arxiv.org/abs/2604.00009,False,core,representation,model,retrieval_access;evaluation_diagnosis;security_privacy_trust,,5,episodic-memory
 2603.11721,2026-03-12,transition-2026-Feb-Mar,When OpenClaw Meets Hospital: Toward an Agentic Operating System for Dynamic Clinical Workflows,https://arxiv.org/abs/2603.11721,False,core,execution_state,model,retrieval_access;representation_organization;experience_skill_learning;evaluation_diagnosis;multimodal_embodied_gui;systems_efficiency,,7,long-term-memory
 2603.11768,2026-03-12,transition-2026-Feb-Mar,"Governing Evolving Memory in LLM Agents: Risks, Mechanisms, and the Stability and Safety Governed Memory (SSGM) Framework",https://arxiv.org/abs/2603.11768,True,core,execution_state,model,retrieval_access;representation_organization;update_forgetting;security_privacy_trust;multimodal_embodied_gui;systems_efficiency,experience_skill_policy;security_governance,0,agent-memory;long-term-memory
 2603.12631,2026-03-13,transition-2026-Feb-Mar,Joint Optimization of Multi-agent Memory System,https://arxiv.org/abs/2603.12631,True,core,update_forgetting,model,retrieval_access;execution_state;experience_skill_learning;multi_agent_shared;multimodal_embodied_gui,personal_shared,5,agent-memory
 2603.13017,2026-03-13,transition-2026-Feb-Mar,Structured Distillation for Personalized Agent Memory: 11x Token Reduction with Retrieval Preservation,https://arxiv.org/abs/2603.13017,True,core,access,model,retrieval_access;representation_organization;evaluation_diagnosis;systems_efficiency,retrieval_representation;experience_skill_policy;systems_efficiency;personal_shared,8,agent-memory
 2603.13676,2026-03-14,transition-2026-Feb-Mar,TheraAgent: Multi-Agent Framework with Self-Evolving Memory and Evidence-Calibrated Reasoning for PET Theranostics,https://arxiv.org/abs/2603.13676,True,core,execution_state,model,representation_organization;experience_skill_learning;evaluation_diagnosis;security_privacy_trust;multi_agent_shared,experience_skill_policy;personal_shared,7,agent-memory
 2603.14588,2026-03-15,transition-2026-Feb-Mar,SuperLocalMemory V3: Information-Geometric Foundations for Zero-LLM Enterprise Agent Memory,https://arxiv.org/abs/2603.14588,True,core,access,model,retrieval_access;representation_organization;evaluation_diagnosis,,6,agent-memory
 2603.14597,2026-03-15,transition-2026-Feb-Mar,D-MEM: Dopamine-Gated Agentic Memory via Reward Prediction Error Routing,https://arxiv.org/abs/2603.14597,True,core,update_forgetting,model,representation_organization;evaluation_diagnosis;security_privacy_trust;systems_efficiency,,7,agent-memory;long-term-memory
 2603.15280,2026-03-16,transition-2026-Feb-Mar,Advancing Multimodal Agent Reasoning with Long-Term Neuro-Symbolic Memory,https://arxiv.org/abs/2603.15280,True,core,execution_state,model,retrieval_access;representation_organization;update_forgetting;experience_skill_learning;evaluation_diagnosis;multimodal_embodied_gui,execution_multimodal,5,agent-memory
 2603.15421,2026-03-16,transition-2026-Feb-Mar,CLAG: Adaptive Memory Organization via Agent-Driven Clustering for Small Language Model Agents,https://arxiv.org/abs/2603.15421,True,core,access,model,retrieval_access;representation_organization;experience_skill_learning;systems_efficiency,retrieval_representation;experience_skill_policy,2,agent-memory
 2603.16734,2026-03-17,transition-2026-Feb-Mar,Differential Harm Propensity in Personalized LLM Agents: The Curious Case of Mental Health Disclosure,https://arxiv.org/abs/2603.16734,False,core,execution_state,model,evaluation_diagnosis;security_privacy_trust,personal_shared,4,long-term-memory
 2603.17043,2026-03-17,transition-2026-Feb-Mar,OpenQlaw: An Agentic AI Assistant for Analysis of 2D Quantum Materials,https://arxiv.org/abs/2603.17043,False,core,execution_state,model,retrieval_access;multimodal_embodied_gui;systems_efficiency,benchmark_diagnosis,0,long-term-memory
 2603.17244,2026-03-18,transition-2026-Feb-Mar,Graph-Native Cognitive Memory for AI Agents: Formal Belief Revision Semantics for Versioned Memory Architectures,https://arxiv.org/abs/2603.17244,True,core,execution_state,model,retrieval_access;representation_organization;evaluation_diagnosis;security_privacy_trust;systems_efficiency,retrieval_representation;update_forgetting,6,agent-memory;memory-evaluation
 2603.17831,2026-03-18,transition-2026-Feb-Mar,RPMS: Enhancing LLM-Based Embodied Planning through Rule-Augmented Memory Synergy,https://arxiv.org/abs/2603.17831,True,core,update_forgetting,model,retrieval_access;representation_organization;update_forgetting;execution_state;evaluation_diagnosis;multimodal_embodied_gui,execution_multimodal,6,episodic-memory
 2603.17948,2026-03-18,transition-2026-Feb-Mar,VideoAtlas: Navigating Long-Form Video in Logarithmic Compute,https://arxiv.org/abs/2603.17948,False,core,experience_learning,model,representation_organization;evaluation_diagnosis;multimodal_embodied_gui;systems_efficiency,execution_multimodal,2,agent-memory
 2603.19935,2026-03-20,transition-2026-Feb-Mar,"Memori: A Persistent Memory Layer for Efficient, Context-Aware LLM Agents",https://arxiv.org/abs/2603.19935,True,core,experience_learning,model,retrieval_access;representation_organization;evaluation_diagnosis;systems_efficiency,systems_efficiency,6,long-term-memory
 2604.14158,2026-03-23,transition-2026-Feb-Mar,MemGround: Long-Term Memory Evaluation Kit for Large Language Models in Gamified Scenarios,https://arxiv.org/abs/2604.14158,True,supporting,evaluation,model,retrieval_access;representation_organization;execution_state;evaluation_diagnosis,benchmark_diagnosis,3,agent-memory
 2605.20189,2026-03-23,transition-2026-Feb-Mar,SOLAR: A Self-Optimizing Open-Ended Autonomous Agent for Lifelong Learning and Continual Adaptation,https://arxiv.org/abs/2605.20189,False,supporting,evaluation,model,update_forgetting;experience_skill_learning;systems_efficiency,experience_skill_policy,6,episodic-memory
 2603.23064,2026-03-24,transition-2026-Feb-Mar,Mind Your HEARTBEAT! Claw Background Execution Inherently Enables Silent Memory Pollution,https://arxiv.org/abs/2603.23064,True,peripheral,security_privacy,model,retrieval_access;evaluation_diagnosis;security_privacy_trust,,3,agent-memory
 2603.23160,2026-03-24,transition-2026-Feb-Mar,UniDial-EvalKit: A Unified Toolkit for Evaluating Multi-Faceted Conversational Abilities,https://arxiv.org/abs/2603.23160,False,supporting,evaluation,model,representation_organization;evaluation_diagnosis;systems_efficiency,benchmark_diagnosis,6,agent-memory
 2603.23231,2026-03-24,transition-2026-Feb-Mar,PERMA: Benchmarking Personalized Memory Agents via Event-Driven Preference and Realistic Task Environments,https://arxiv.org/abs/2603.23231,True,core,experience_learning,model,retrieval_access;evaluation_diagnosis;multimodal_embodied_gui;systems_efficiency,benchmark_diagnosis;personal_shared,5,agent-memory
 2603.24564,2026-03-25,transition-2026-Feb-Mar,"Infrastructure for Valuable, Tradable, and Verifiable Agent Memory",https://arxiv.org/abs/2603.24564,True,supporting,evaluation,model,representation_organization;security_privacy_trust;systems_efficiency,retrieval_representation,0,agent-memory
 2603.25097,2026-03-26,transition-2026-Feb-Mar,ElephantBroker: A Knowledge-Grounded Cognitive Runtime for Trustworthy AI Agents,https://arxiv.org/abs/2603.25097,False,core,evaluation,model,retrieval_access;representation_organization;update_forgetting;security_privacy_trust;systems_efficiency,security_governance;systems_efficiency,0,agent-memory
 2603.25973,2026-03-26,transition-2026-Feb-Mar,MemoryCD: Benchmarking Long-Context User Memory of LLM Agents for Lifelong Cross-Domain Personalization,https://arxiv.org/abs/2603.25973,True,supporting,evaluation,model,evaluation_diagnosis;systems_efficiency,benchmark_diagnosis;personal_shared,5,memory-evaluation
 2604.19771,2026-03-27,transition-2026-Feb-Mar,Cognis: Context-Aware Memory for Conversational AI Agents,https://arxiv.org/abs/2604.19771,True,core,experience_learning,model,retrieval_access;evaluation_diagnosis;systems_efficiency,,4,long-term-memory
 2603.28088,2026-03-30,transition-2026-Feb-Mar,GEMS: Agent-Native Multimodal Generation with Memory and Skills,https://arxiv.org/abs/2603.28088,True,core,experience_learning,model,representation_organization;execution_state;experience_skill_learning;evaluation_diagnosis;multi_agent_shared;multimodal_embodied_gui,experience_skill_policy;execution_multimodal,4,agent-memory
 2603.29493,2026-03-31,transition-2026-Feb-Mar,MemFactory: Unified Inference & Training Framework for Agent Memory,https://arxiv.org/abs/2603.29493,True,core,update_forgetting,model,retrieval_access;representation_organization;experience_skill_learning;evaluation_diagnosis,,6,agent-memory
 2604.00131,2026-03-31,transition-2026-Feb-Mar,Oblivion: Self-Adaptive Agentic Memory Control through Decay-Driven Activation,https://arxiv.org/abs/2604.00131,True,core,update_forgetting,model,retrieval_access;representation_organization;update_forgetting;execution_state;experience_skill_learning;evaluation_diagnosis;systems_efficiency,update_forgetting;experience_skill_policy,2,agent-memory
 2604.00556,2026-04-01,recent-2026-Apr-Jul,HabitatAgent: An End-to-End Multi-Agent System for Housing Consultation,https://arxiv.org/abs/2604.00556,False,core,experience_learning,model,retrieval_access;update_forgetting;evaluation_diagnosis;multi_agent_shared,personal_shared,6,agent-memory
 2604.01007,2026-04-01,recent-2026-Apr-Jul,Omni-SimpleMem: Autoresearch-Guided Discovery of Lifelong Multimodal Agent Memory,https://arxiv.org/abs/2604.01007,True,core,update_forgetting,model,retrieval_access;experience_skill_learning;evaluation_diagnosis;multimodal_embodied_gui,retrieval_representation;execution_multimodal,8,agent-memory
 2604.01560,2026-04-02,recent-2026-Apr-Jul,DeltaMem: Towards Agentic Memory Management via Reinforcement Learning,https://arxiv.org/abs/2604.01560,True,core,update_forgetting,model,experience_skill_learning;evaluation_diagnosis;multi_agent_shared,experience_skill_policy,4,agent-memory;memory-evaluation
 2604.01658,2026-04-02,recent-2026-Apr-Jul,CORAL: Towards Autonomous Multi-Agent Evolution for Open-Ended Discovery,https://arxiv.org/abs/2604.01658,False,core,update_forgetting,model,retrieval_access;experience_skill_learning;evaluation_diagnosis;multi_agent_shared,personal_shared,5,long-term-memory
 2604.01707,2026-04-02,recent-2026-Apr-Jul,Memory in the LLM Era: Modular Architectures and Strategies in a Unified Framework,https://arxiv.org/abs/2604.01707,True,core,access,model,retrieval_access;execution_state;evaluation_diagnosis,,4,agent-memory
 2604.02522,2026-04-02,recent-2026-Apr-Jul,Opal: Private Memory for Personal AI,https://arxiv.org/abs/2604.02522,True,core,access,model,retrieval_access;representation_organization;evaluation_diagnosis;security_privacy_trust;systems_efficiency,,6,agent-memory
 2604.02623,2026-04-03,recent-2026-Apr-Jul,"Poison Once, Exploit Forever: Environment-Injected Memory Poisoning Attacks on Web Agents",https://arxiv.org/abs/2604.02623,True,core,update_forgetting,model,retrieval_access;execution_state;security_privacy_trust;multi_agent_shared;multimodal_embodied_gui;systems_efficiency,execution_multimodal;security_governance,4,agent-memory
 2604.03588,2026-04-04,recent-2026-Apr-Jul,Rashomon Memory: Towards Argumentation-Driven Retrieval for Multi-Perspective Agent Memory,https://arxiv.org/abs/2604.03588,True,core,experience_learning,model,retrieval_access;representation_organization;update_forgetting;experience_skill_learning;evaluation_diagnosis;security_privacy_trust;systems_efficiency,retrieval_representation,0,agent-memory
 2604.04157,2026-04-05,recent-2026-Apr-Jul,Readable Minds: Emergent Theory-of-Mind-Like Behavior in LLM Poker Agents,https://arxiv.org/abs/2604.04157,False,core,execution_state,model,,,8,long-term-memory
 2604.04503,2026-04-06,recent-2026-Apr-Jul,Memory Intelligence Agent,https://arxiv.org/abs/2604.04503,True,core,update_forgetting,model,retrieval_access;update_forgetting;experience_skill_learning;evaluation_diagnosis;multimodal_embodied_gui;systems_efficiency,,3,agent-memory
 2604.04514,2026-04-06,recent-2026-Apr-Jul,"SuperLocalMemory V3.3: The Living Brain -- Biologically-Inspired Forgetting, Cognitive Quantization, and Multi-Channel Retrieval for Zero-LLM Agent Memory Systems",https://arxiv.org/abs/2604.04514,True,core,experience_learning,model,retrieval_access;representation_organization;update_forgetting;experience_skill_learning;evaluation_diagnosis;security_privacy_trust;systems_efficiency,retrieval_representation;update_forgetting,4,agent-memory
 2604.04660,2026-04-06,recent-2026-Apr-Jul,"Springdrift: An Auditable Persistent Runtime for LLM Agents with Case-Based Memory, Normative Safety, and Ambient Self-Perception",https://arxiv.org/abs/2604.04660,True,supporting,execution_state,model,retrieval_access;representation_organization;evaluation_diagnosis;multimodal_embodied_gui,systems_efficiency,6,long-term-memory
 2604.04853,2026-04-06,recent-2026-Apr-Jul,MemMachine: A Ground-Truth-Preserving Memory System for Personalized AI Agents,https://arxiv.org/abs/2604.04853,True,core,experience_learning,model,retrieval_access;execution_state;evaluation_diagnosis;systems_efficiency,systems_efficiency;personal_shared,6,long-term-memory
 2604.04901,2026-04-06,recent-2026-Apr-Jul,FileGram: Grounding Agent Personalization in File-System Behavioral Traces,https://arxiv.org/abs/2604.04901,False,supporting,execution_state,model,retrieval_access;experience_skill_learning;evaluation_diagnosis;security_privacy_trust;multimodal_embodied_gui,personal_shared,2,agent-memory
 2604.05719,2026-04-07,recent-2026-Apr-Jul,Hackers or Hallucinators? A Comprehensive Analysis of LLM-Based Automated Penetration Testing,https://arxiv.org/abs/2604.05719,False,supporting,execution_state,model,retrieval_access;representation_organization;evaluation_diagnosis;security_privacy_trust;systems_efficiency,benchmark_diagnosis,5,agent-memory
 2604.07798,2026-04-09,recent-2026-Apr-Jul,Lightweight LLM Agent Memory with Small Language Models,https://arxiv.org/abs/2604.07798,True,core,experience_learning,model,retrieval_access;update_forgetting;execution_state;systems_efficiency,,2,agent-memory;long-term-memory
 2604.07877,2026-04-09,recent-2026-Apr-Jul,MemReader: From Passive to Active Extraction for Long-Term Agent Memory,https://arxiv.org/abs/2604.07877,True,core,experience_learning,model,retrieval_access;representation_organization;evaluation_diagnosis;multimodal_embodied_gui;systems_efficiency,,5,agent-memory
 2604.08064,2026-04-09,recent-2026-Apr-Jul,ImplicitMemBench: Measuring Unconscious Behavioral Adaptation in Large Language Models,https://arxiv.org/abs/2604.08064,False,supporting,execution_state,model,retrieval_access;experience_skill_learning;evaluation_diagnosis,experience_skill_policy,8,memory-evaluation;procedural-memory
 2604.08756,2026-04-09,recent-2026-Apr-Jul,Artifacts as Memory Beyond the Agent Boundary,https://arxiv.org/abs/2604.08756,True,supporting,execution_state,model,experience_skill_learning,,2,agent-memory
 2604.09000,2026-04-10,recent-2026-Apr-Jul,StreamMeCo: Long-Term Agent Memory Compression for Efficient Streaming Video Understanding,https://arxiv.org/abs/2604.09000,True,core,experience_learning,model,retrieval_access;evaluation_diagnosis;multimodal_embodied_gui;systems_efficiency,execution_multimodal;systems_efficiency,5,agent-memory
 2604.09388,2026-04-10,recent-2026-Apr-Jul,The AI Codebase Maturity Model: From Assisted Coding to Fully Autonomous Systems,https://arxiv.org/abs/2604.09388,False,core,experience_learning,model,representation_organization;experience_skill_learning;evaluation_diagnosis;multi_agent_shared;systems_efficiency,,3,agent-memory
 2604.09747,2026-04-10,recent-2026-Apr-Jul,ADAM: A Systematic Data Extraction Attack on Agent Memory via Adaptive Querying,https://arxiv.org/abs/2604.09747,True,supporting,execution_state,model,retrieval_access;security_privacy_trust;multimodal_embodied_gui;systems_efficiency,experience_skill_policy;security_governance,6,agent-memory
 2604.11811,2026-04-10,recent-2026-Apr-Jul,M$^\star$: Every Task Deserves Its Own Memory Harness,https://arxiv.org/abs/2604.11811,True,core,update_forgetting,model,retrieval_access;execution_state;experience_skill_learning;evaluation_diagnosis;multimodal_embodied_gui;systems_efficiency,,5,agent-memory
 2604.10981,2026-04-13,recent-2026-Apr-Jul,"ATANT v1.1: Positioning Continuity Evaluation Against Memory, Long-Context, and Agentic-Memory Benchmarks",https://arxiv.org/abs/2604.10981,True,supporting,evaluation,model,evaluation_diagnosis;security_privacy_trust,benchmark_diagnosis,5,agent-memory;memory-evaluation
 2604.11544,2026-04-13,recent-2026-Apr-Jul,Time is Not a Label: Continuous Phase Rotation for Temporal Knowledge Graphs and Agentic Memory,https://arxiv.org/abs/2604.11544,True,core,update_forgetting,model,representation_organization;evaluation_diagnosis;multimodal_embodied_gui,retrieval_representation;systems_efficiency,4,agent-memory
 2604.11563,2026-04-13,recent-2026-Apr-Jul,Synthius-Mem: Brain-Inspired Hallucination-Resistant Persona Memory Achieving 94.4% Memory Accuracy and 99.6% Adversarial Robustness on LoCoMo,https://arxiv.org/abs/2604.11563,True,core,update_forgetting,model,retrieval_access;representation_organization;update_forgetting;experience_skill_learning;evaluation_diagnosis;security_privacy_trust;systems_efficiency,,8,long-term-memory
 2604.12007,2026-04-13,recent-2026-Apr-Jul,When to Forget: A Memory Governance Primitive,https://arxiv.org/abs/2604.12007,True,core,update_forgetting,model,retrieval_access;update_forgetting;experience_skill_learning;evaluation_diagnosis;security_privacy_trust,update_forgetting;security_governance,6,agent-memory
 2604.12179,2026-04-14,recent-2026-Apr-Jul,AgenticAI-DialogGen: Topic-Guided Conversation Generation for Fine-Tuning and Evaluating Short- and Long-Term Memories of LLMs,https://arxiv.org/abs/2604.12179,True,core,update_forgetting,model,representation_organization;evaluation_diagnosis;multimodal_embodied_gui;systems_efficiency,benchmark_diagnosis;execution_multimodal,2,long-term-memory
 2604.12285,2026-04-14,recent-2026-Apr-Jul,GAM: Hierarchical Graph-based Agentic Memory for LLM Agents,https://arxiv.org/abs/2604.12285,True,core,update_forgetting,model,retrieval_access;representation_organization;update_forgetting;multimodal_embodied_gui;systems_efficiency,retrieval_representation,4,agent-memory
 2604.12948,2026-04-14,recent-2026-Apr-Jul,Drawing on Memory: Dual-Trace Encoding Improves Cross-Session Recall in LLM Agents,https://arxiv.org/abs/2604.12948,True,core,update_forgetting,model,retrieval_access;update_forgetting;evaluation_diagnosis;systems_efficiency,retrieval_representation,7,long-term-memory
 2604.15774,2026-04-17,recent-2026-Apr-Jul,MemEvoBench: Benchmarking Safety Risks from Memory Misevolution in LLM Agents,https://arxiv.org/abs/2604.15774,True,supporting,evaluation,model,update_forgetting;execution_state;evaluation_diagnosis;security_privacy_trust,benchmark_diagnosis,4,long-term-memory
 2604.15877,2026-04-17,recent-2026-Apr-Jul,"Experience Compression Spectrum: Unifying Memory, Skills, and Rules in LLM Agents",https://arxiv.org/abs/2604.15877,True,supporting,evaluation,model,retrieval_access;execution_state;experience_skill_learning;evaluation_diagnosis;systems_efficiency,experience_skill_policy;systems_efficiency,2,agent-memory;episodic-memory
 2604.16548,2026-04-17,recent-2026-Apr-Jul,"A Survey on Long-Term Memory Security in LLM Agents: Attacks, Defenses, and Governance Across the Memory Lifecycle",https://arxiv.org/abs/2604.16548,True,core,update_forgetting,model,retrieval_access;update_forgetting;security_privacy_trust;systems_efficiency,benchmark_diagnosis;update_forgetting;security_governance,0,long-term-memory
 2604.16839,2026-04-18,recent-2026-Apr-Jul,HeLa-Mem: Hebbian Learning and Associative Memory for LLM Agents,https://arxiv.org/abs/2604.16839,True,core,update_forgetting,model,retrieval_access;representation_organization;update_forgetting;experience_skill_learning;systems_efficiency,experience_skill_policy,5,episodic-memory;long-term-memory
 2604.17273,2026-04-19,recent-2026-Apr-Jul,The Continuity Layer: Why Intelligence Needs an Architecture for What It Carries Forward,https://arxiv.org/abs/2604.17273,False,core,execution_state,model,retrieval_access;representation_organization;execution_state;evaluation_diagnosis;security_privacy_trust;systems_efficiency,,3,agent-memory;memory-evaluation
 2604.17456,2026-04-19,recent-2026-Apr-Jul,TrafficClaw: A Generalizable LLM Agent in the Unified Physical Environment for Urban Traffic Control,https://arxiv.org/abs/2604.17456,False,core,execution_state,model,representation_organization;execution_state;experience_skill_learning,,3,long-term-memory
 2604.17658,2026-04-19,recent-2026-Apr-Jul,Towards Self-Improving Error Diagnosis in Multi-Agent Systems,https://arxiv.org/abs/2604.17658,False,core,execution_state,model,update_forgetting;evaluation_diagnosis;multi_agent_shared,benchmark_diagnosis;personal_shared,6,episodic-memory
 2604.19541,2026-04-21,recent-2026-Apr-Jul,FOCAL: Filtered On-device Continuous Activity Logging for Efficient Personal Desktop Summarization,https://arxiv.org/abs/2604.19541,False,core,execution_state,model,retrieval_access;security_privacy_trust;multi_agent_shared;multimodal_embodied_gui;systems_efficiency,systems_efficiency,6,agent-memory
 2604.20006,2026-04-21,recent-2026-Apr-Jul,From Recall to Forgetting: Benchmarking Long-Term Memory for Personalized Agents,https://arxiv.org/abs/2604.20006,True,core,execution_state,model,retrieval_access;update_forgetting;evaluation_diagnosis,benchmark_diagnosis;retrieval_representation;update_forgetting;personal_shared,3,agent-memory
 2604.20117,2026-04-22,recent-2026-Apr-Jul,To Know is to Construct: Schema-Constrained Generation for Agent Memory,https://arxiv.org/abs/2604.20117,True,core,execution_state,model,retrieval_access;representation_organization;update_forgetting;evaluation_diagnosis;multimodal_embodied_gui,,5,agent-memory
 2604.20183,2026-04-22,recent-2026-Apr-Jul,Dual-Cluster Memory Agent: Resolving Multi-Paradigm Ambiguity in Optimization Problem Solving,https://arxiv.org/abs/2604.20183,True,core,execution_state,model,representation_organization;update_forgetting;evaluation_diagnosis;multimodal_embodied_gui,execution_multimodal,4,agent-memory
 2604.20300,2026-04-22,recent-2026-Apr-Jul,FSFM: A Biologically-Inspired Framework for Selective Forgetting of Agent Memory,https://arxiv.org/abs/2604.20300,True,core,execution_state,model,retrieval_access;representation_organization;update_forgetting;evaluation_diagnosis;security_privacy_trust,update_forgetting,5,agent-memory
 2604.20582,2026-04-22,recent-2026-Apr-Jul,"Trust, Lies, and Long Memories: Emergent Social Dynamics and Reputation in Multi-Round Avalon with LLM Agents",https://arxiv.org/abs/2604.20582,True,core,execution_state,model,security_privacy_trust,security_governance,4,agent-memory
 2604.20598,2026-04-22,recent-2026-Apr-Jul,"Self-Aware Vector Embeddings for Retrieval-Augmented Generation: A Neuroscience-Inspired Framework for Temporal, Confidence-Weighted, and Relational Knowledge",https://arxiv.org/abs/2604.20598,False,core,execution_state,model,retrieval_access;representation_organization;update_forgetting;evaluation_diagnosis;security_privacy_trust;systems_efficiency,retrieval_representation;systems_efficiency,6,agent-memory
 2604.22085,2026-04-23,recent-2026-Apr-Jul,Memanto: Typed Semantic Memory with Information-Theoretic Retrieval for Long-Horizon Agents,https://arxiv.org/abs/2604.22085,True,core,execution_state,model,retrieval_access;representation_organization;update_forgetting;execution_state;evaluation_diagnosis;systems_efficiency,retrieval_representation;execution_multimodal,3,agent-memory
 2604.23711,2026-04-26,recent-2026-Apr-Jul,Spore: Efficient and Training-Free Privacy Extraction Attack on LLMs via Inference-Time Hybrid Probing,https://arxiv.org/abs/2604.23711,False,core,execution_state,model,retrieval_access;evaluation_diagnosis;security_privacy_trust;systems_efficiency,security_governance;systems_efficiency,4,agent-memory
 2606.20570,2026-04-26,recent-2026-Apr-Jul,Infrastructure for the Agentic Web: Gap Analysis and Architecture from the Agentverse Platform,https://arxiv.org/abs/2606.20570,False,core,access,model,retrieval_access;representation_organization;evaluation_diagnosis;security_privacy_trust;multimodal_embodied_gui;systems_efficiency,benchmark_diagnosis;retrieval_representation,1,agent-memory
 2604.26197,2026-04-29,recent-2026-Apr-Jul,Hierarchical Long-Term Semantic Memory for LinkedIn's Hiring Agent,https://arxiv.org/abs/2604.26197,True,core,execution_state,model,retrieval_access;representation_organization;evaluation_diagnosis;security_privacy_trust;systems_efficiency,retrieval_representation,5,long-term-memory
 2604.26622,2026-04-29,recent-2026-Apr-Jul,OCR-Memory: Optical Context Retrieval for Long-Horizon Agent Memory,https://arxiv.org/abs/2604.26622,True,core,update_forgetting,model,retrieval_access;representation_organization;update_forgetting;execution_state;experience_skill_learning;evaluation_diagnosis;multimodal_embodied_gui;systems_efficiency,retrieval_representation;execution_multimodal,3,agent-memory
 2604.27003,2026-04-29,recent-2026-Apr-Jul,When Continual Learning Moves to Memory: A Study of Experience Reuse in LLM Agents,https://arxiv.org/abs/2604.27003,True,core,update_forgetting,model,retrieval_access;representation_organization;update_forgetting;experience_skill_learning;evaluation_diagnosis,benchmark_diagnosis;experience_skill_policy,3,procedural-memory
 2604.27045,2026-04-29,recent-2026-Apr-Jul,Detecting Clinical Discrepancies in Health Coaching Agents: A Dual-Stream Memory and Reconciliation Architecture,https://arxiv.org/abs/2604.27045,True,core,update_forgetting,model,retrieval_access;representation_organization;update_forgetting;evaluation_diagnosis;security_privacy_trust,,5,agent-memory
 2604.27283,2026-04-30,recent-2026-Apr-Jul,Learning When to Remember: Risk-Sensitive Contextual Bandits for Abstention-Aware Memory Retrieval in LLM-Based Coding Agents,https://arxiv.org/abs/2604.27283,True,core,update_forgetting,model,retrieval_access;execution_state;experience_skill_learning;systems_efficiency,retrieval_representation;experience_skill_policy,3,agent-memory
 2604.27707,2026-04-30,recent-2026-Apr-Jul,"Contextual Agentic Memory is a Memo, Not True Memory",https://arxiv.org/abs/2604.27707,True,core,update_forgetting,model,retrieval_access;update_forgetting;evaluation_diagnosis;security_privacy_trust;systems_efficiency,,3,agent-memory
 2604.27996,2026-04-30,recent-2026-Apr-Jul,Exploring LLM Agent Designs and Interaction Modalities for Scientific Visualization,https://arxiv.org/abs/2604.27996,False,core,update_forgetting,model,representation_organization;evaluation_diagnosis;multimodal_embodied_gui;systems_efficiency,,4,long-term-memory
 2605.00702,2026-05-01,recent-2026-Apr-Jul,Learning How and What to Memorize: Cognition-Inspired Two-Stage Optimization for Evolving Memory,https://arxiv.org/abs/2605.00702,True,core,update_forgetting,model,representation_organization;update_forgetting;execution_state;experience_skill_learning;evaluation_diagnosis;multimodal_embodied_gui,experience_skill_policy,4,memory-evaluation
 2605.01386,2026-05-02,recent-2026-Apr-Jul,MemORAI: Memory Organization and Retrieval via Adaptive Graph Intelligence for LLM Conversational Agents,https://arxiv.org/abs/2605.01386,True,core,representation,model,retrieval_access;representation_organization;evaluation_diagnosis;security_privacy_trust;systems_efficiency,retrieval_representation;experience_skill_policy,2,long-term-memory
 2605.01970,2026-05-03,recent-2026-Apr-Jul,Trojan Hippo: Weaponizing Agent Memory for Data Exfiltration,https://arxiv.org/abs/2605.01970,True,core,update_forgetting,model,evaluation_diagnosis;security_privacy_trust;systems_efficiency,,4,agent-memory;long-term-memory
 2605.02199,2026-05-04,recent-2026-Apr-Jul,MEMAUDIT: An Exact Package-Oracle Evaluation Protocol for Budgeted Long-Term LLM Memory Writing,https://arxiv.org/abs/2605.02199,True,core,update_forgetting,model,retrieval_access;representation_organization;experience_skill_learning;evaluation_diagnosis;systems_efficiency,benchmark_diagnosis;systems_efficiency,3,long-term-memory
 2605.03228,2026-05-04,recent-2026-Apr-Jul,MAGE: Safeguarding LLM Agents against Long-Horizon Threats via Shadow Memory,https://arxiv.org/abs/2605.03228,True,core,update_forgetting,model,retrieval_access;execution_state;evaluation_diagnosis;security_privacy_trust,execution_multimodal,5,agent-memory
 2605.03312,2026-05-05,recent-2026-Apr-Jul,MemFlow: Intent-Driven Memory Orchestration for Small Language Model Agents,https://arxiv.org/abs/2605.03312,True,core,update_forgetting,model,retrieval_access;representation_organization;execution_state;evaluation_diagnosis;systems_efficiency,,7,agent-memory;memory-evaluation
 2605.03354,2026-05-05,recent-2026-Apr-Jul,What Happens Inside Agent Memory? Circuit Analysis from Emergence to Diagnosis,https://arxiv.org/abs/2605.03354,True,core,execution_state,model,retrieval_access;representation_organization;evaluation_diagnosis;multimodal_embodied_gui,benchmark_diagnosis,5,agent-memory
 2605.03804,2026-05-05,recent-2026-Apr-Jul,ScrapMem: A Bio-inspired Framework for On-device Personalized Agent Memory via Optical Forgetting,https://arxiv.org/abs/2605.03804,True,core,update_forgetting,model,retrieval_access;representation_organization;update_forgetting;multimodal_embodied_gui;systems_efficiency,update_forgetting;systems_efficiency;personal_shared,4,agent-memory;episodic-memory;long-term-memory
 2605.06702,2026-05-05,recent-2026-Apr-Jul,CASCADE: Case-Based Continual Adaptation for Large Language Models During Deployment,https://arxiv.org/abs/2605.06702,False,core,update_forgetting,model,retrieval_access;experience_skill_learning;evaluation_diagnosis;multimodal_embodied_gui,experience_skill_policy,5,episodic-memory
 2605.04811,2026-05-06,recent-2026-Apr-Jul,Tree-based Credit Assignment for Multi-Agent Memory System,https://arxiv.org/abs/2605.04811,True,core,update_forgetting,model,retrieval_access;representation_organization;update_forgetting;execution_state;experience_skill_learning;evaluation_diagnosis;multi_agent_shared;systems_efficiency,experience_skill_policy;personal_shared,5,agent-memory
 2605.04897,2026-05-06,recent-2026-Apr-Jul,Storage Is Not Memory: A Retrieval-Centered Architecture for Agent Recall,https://arxiv.org/abs/2605.04897,True,core,access,model,retrieval_access;representation_organization;evaluation_diagnosis;systems_efficiency,retrieval_representation;systems_efficiency,3,agent-memory
 2605.05583,2026-05-07,recent-2026-Apr-Jul,Belief Memory: Agent Memory Under Partial Observability,https://arxiv.org/abs/2605.05583,True,core,update_forgetting,model,retrieval_access;update_forgetting;evaluation_diagnosis,,5,agent-memory
 2605.06132,2026-05-07,recent-2026-Apr-Jul,MemReranker: Reasoning-Aware Reranking for Agent Memory Retrieval,https://arxiv.org/abs/2605.06132,True,core,update_forgetting,model,retrieval_access;evaluation_diagnosis;systems_efficiency,retrieval_representation,7,agent-memory
 2605.06527,2026-05-07,recent-2026-Apr-Jul,STALE: Can LLM Agents Know When Their Memories Are No Longer Valid?,https://arxiv.org/abs/2605.06527,True,core,update_forgetting,model,retrieval_access;representation_organization;update_forgetting;evaluation_diagnosis;systems_efficiency,update_forgetting,7,agent-memory
 2605.06716,2026-05-07,recent-2026-Apr-Jul,From Storage to Experience: A Survey on the Evolution of LLM Agent Memory Mechanisms,https://arxiv.org/abs/2605.06716,True,core,experience_learning,model,retrieval_access;execution_state;experience_skill_learning;systems_efficiency,benchmark_diagnosis;experience_skill_policy;systems_efficiency,0,agent-memory
 2605.06812,2026-05-07,recent-2026-Apr-Jul,Towards Security-Auditable LLM Agents: A Unified Graph Representation,https://arxiv.org/abs/2605.06812,False,core,security_privacy,model,retrieval_access;representation_organization;evaluation_diagnosis;security_privacy_trust;multi_agent_shared,retrieval_representation;security_governance,3,long-term-memory
 2605.07242,2026-05-08,recent-2026-Apr-Jul,MEMOREPAIR: Barrier-First Cascade Repair in Agentic Memory,https://arxiv.org/abs/2605.07242,True,core,update_forgetting,model,update_forgetting;experience_skill_learning;evaluation_diagnosis;security_privacy_trust;systems_efficiency,,8,agent-memory
 2605.07313,2026-05-08,recent-2026-Apr-Jul,When Stored Evidence Stops Being Usable: Scale-Conditioned Evaluation of Agent Memory,https://arxiv.org/abs/2605.07313,True,supporting,evaluation,model,retrieval_access;representation_organization;evaluation_diagnosis;systems_efficiency,benchmark_diagnosis,3,agent-memory
 2605.08374,2026-05-08,recent-2026-Apr-Jul,MemQ: Integrating Q-Learning into Self-Evolving Memory Agents over Provenance DAGs,https://arxiv.org/abs/2605.08374,True,core,update_forgetting,model,retrieval_access;representation_organization;update_forgetting;experience_skill_learning;evaluation_diagnosis;security_privacy_trust;multimodal_embodied_gui,experience_skill_policy;security_governance,3,agent-memory;episodic-memory
 2605.08442,2026-05-08,recent-2026-Apr-Jul,Defense effectiveness across architectural layers: a mechanistic evaluation of persistent memory attacks on stateful LLM agents,https://arxiv.org/abs/2605.08442,True,supporting,security_privacy,model,retrieval_access;evaluation_diagnosis;security_privacy_trust;multimodal_embodied_gui;systems_efficiency,benchmark_diagnosis;security_governance,7,long-term-memory
 2605.08468,2026-05-08,recent-2026-Apr-Jul,"PYTHALAB-MERA: Validation-Grounded Memory, Retrieval, and Acceptance Control for Frozen-LLM Coding Agents",https://arxiv.org/abs/2605.08468,True,core,update_forgetting,model,retrieval_access;experience_skill_learning;evaluation_diagnosis;systems_efficiency,retrieval_representation,5,episodic-memory
 2605.08538,2026-05-08,recent-2026-Apr-Jul,Human-Inspired Memory Architecture for LLM Agents,https://arxiv.org/abs/2605.08538,True,core,update_forgetting,model,retrieval_access;representation_organization;update_forgetting;evaluation_diagnosis;security_privacy_trust;systems_efficiency,,8,long-term-memory
 2605.09033,2026-05-09,recent-2026-Apr-Jul,ShadowMerge: A Novel Poisoning Attack on Graph-Based Agent Memory via Relation-Channel Conflicts,https://arxiv.org/abs/2605.09033,True,core,update_forgetting,model,retrieval_access;representation_organization;update_forgetting;evaluation_diagnosis;security_privacy_trust,retrieval_representation;security_governance,7,agent-memory
 2605.09315,2026-05-10,recent-2026-Apr-Jul,Do Self-Evolving Agents Forget? Capability Degradation and Preservation in Lifelong LLM Agent Adaptation,https://arxiv.org/abs/2605.09315,True,core,update_forgetting,model,update_forgetting;execution_state;experience_skill_learning;systems_efficiency,update_forgetting;experience_skill_policy,0,long-term-memory
 2605.09330,2026-05-10,recent-2026-Apr-Jul,The Trap of Trajectory: Towards Understanding and Mitigating Spurious Correlations in Agentic Memory,https://arxiv.org/abs/2605.09330,True,supporting,evaluation,model,retrieval_access;representation_organization;execution_state;evaluation_diagnosis;security_privacy_trust;systems_efficiency,,2,agent-memory
 2605.11032,2026-05-10,recent-2026-Apr-Jul,Portable Agent Memory: A Protocol for Cryptographically-Verified Memory Transfer Across Heterogeneous AI Agents,https://arxiv.org/abs/2605.11032,True,core,update_forgetting,model,retrieval_access;representation_organization;experience_skill_learning;security_privacy_trust;systems_efficiency,retrieval_representation,1,agent-memory
 2605.09863,2026-05-11,recent-2026-Apr-Jul,Nautilus Compass: Black-box Persona Drift Detection for Production LLM Agents,https://arxiv.org/abs/2605.09863,False,core,update_forgetting,model,retrieval_access;representation_organization;update_forgetting;systems_efficiency,,7,agent-memory
 2605.09942,2026-05-11,recent-2026-Apr-Jul,HAGE: Harnessing Agentic Memory via RL-Driven Weighted Graph Evolution,https://arxiv.org/abs/2605.09942,True,core,update_forgetting,model,retrieval_access;representation_organization;execution_state;experience_skill_learning;multi_agent_shared,retrieval_representation;experience_skill_policy,3,agent-memory
 2605.10268,2026-05-11,recent-2026-Apr-Jul,MemReread: Enhancing Agentic Long-Context Reasoning via Memory-Guided Rereading,https://arxiv.org/abs/2605.10268,True,core,update_forgetting,model,retrieval_access;update_forgetting;experience_skill_learning;multimodal_embodied_gui;systems_efficiency,execution_multimodal,5,agent-memory
 2605.10870,2026-05-11,recent-2026-Apr-Jul,"Remember the Decision, Not the Description: A Rate-Distortion Framework for Agent Memory",https://arxiv.org/abs/2605.10870,True,core,update_forgetting,model,update_forgetting;execution_state;experience_skill_learning;evaluation_diagnosis;systems_efficiency,,3,agent-memory
 2605.10899,2026-05-11,recent-2026-Apr-Jul,RubricEM: Meta-RL with Rubric-guided Policy Decomposition beyond Verifiable Rewards,https://arxiv.org/abs/2605.10899,False,core,update_forgetting,model,retrieval_access;representation_organization;execution_state;experience_skill_learning;evaluation_diagnosis;multimodal_embodied_gui,experience_skill_policy;execution_multimodal,4,agent-memory
 2605.13880,2026-05-11,recent-2026-Apr-Jul,PREPING: Building Agent Memory without Tasks,https://arxiv.org/abs/2605.13880,True,core,update_forgetting,model,representation_organization;update_forgetting;experience_skill_learning;multimodal_embodied_gui;systems_efficiency,,5,agent-memory
 2605.11814,2026-05-12,recent-2026-Apr-Jul,MedMemoryBench: Benchmarking Agent Memory in Personalized Healthcare,https://arxiv.org/abs/2605.11814,True,core,execution_state,model,retrieval_access;execution_state;evaluation_diagnosis;systems_efficiency,benchmark_diagnosis;personal_shared,3,agent-memory
 2605.12061,2026-05-12,recent-2026-Apr-Jul,SAGE: A Self-Evolving Agentic Graph-Memory Engine for Structure-Aware Associative Memory,https://arxiv.org/abs/2605.12061,True,core,update_forgetting,model,retrieval_access;representation_organization;execution_state;experience_skill_learning;evaluation_diagnosis,retrieval_representation;experience_skill_policy,2,agent-memory;long-term-memory;memory-evaluation
 2605.12213,2026-05-12,recent-2026-Apr-Jul,Goal-Oriented Reasoning for RAG-based Memory in Conversational Agentic LLM Systems,https://arxiv.org/abs/2605.12213,True,core,update_forgetting,model,retrieval_access;execution_state,retrieval_representation,5,agent-memory
 2605.12294,2026-05-12,recent-2026-Apr-Jul,Executable Agentic Memory for GUI Agent,https://arxiv.org/abs/2605.12294,True,core,execution_state,model,retrieval_access;representation_organization;execution_state;multimodal_embodied_gui;systems_efficiency,execution_multimodal,2,agent-memory
 2605.12493,2026-05-12,recent-2026-Apr-Jul,LongMemEval-V2: Evaluating Long-Term Agent Memory Toward Experienced Colleagues,https://arxiv.org/abs/2605.12493,True,core,experience_learning,model,retrieval_access;experience_skill_learning;evaluation_diagnosis;multimodal_embodied_gui;systems_efficiency,benchmark_diagnosis;experience_skill_policy,8,agent-memory;memory-evaluation
 2605.12978,2026-05-13,recent-2026-Apr-Jul,Useful Memories Become Faulty When Continuously Updated by LLMs,https://arxiv.org/abs/2605.12978,True,core,update_forgetting,model,update_forgetting;experience_skill_learning;evaluation_diagnosis,update_forgetting,6,agent-memory
 2605.13438,2026-05-13,recent-2026-Apr-Jul,CogniFold: Always-On Proactive Memory via Cognitive Folding,https://arxiv.org/abs/2605.13438,True,core,representation,model,retrieval_access;representation_organization;update_forgetting;experience_skill_learning;evaluation_diagnosis,,3,agent-memory
 2605.13542,2026-05-13,recent-2026-Apr-Jul,RealICU: Do LLM Agents Understand Long-Context ICU Data? A Benchmark Beyond Behavior Imitation,https://arxiv.org/abs/2605.13542,False,core,execution_state,model,retrieval_access;representation_organization;execution_state;evaluation_diagnosis,benchmark_diagnosis,4,agent-memory
 2605.13941,2026-05-13,recent-2026-Apr-Jul,EvolveMem:Self-Evolving Memory Architecture via AutoResearch for LLM Agents,https://arxiv.org/abs/2605.13941,True,core,update_forgetting,model,retrieval_access;representation_organization;experience_skill_learning;evaluation_diagnosis,retrieval_representation;experience_skill_policy,8,long-term-memory
 2605.14421,2026-05-14,recent-2026-Apr-Jul,MemLineage: Lineage-Guided Enforcement for LLM Agent Memory,https://arxiv.org/abs/2605.14421,True,supporting,security_privacy,model,retrieval_access;evaluation_diagnosis;security_privacy_trust;multimodal_embodied_gui,execution_multimodal,5,agent-memory
 2605.14498,2026-05-14,recent-2026-Apr-Jul,GroupMemBench: Benchmarking LLM Agent Memory in Multi-Party Conversations,https://arxiv.org/abs/2605.14498,True,supporting,evaluation,model,retrieval_access;representation_organization;update_forgetting;experience_skill_learning;evaluation_diagnosis;security_privacy_trust;multimodal_embodied_gui,benchmark_diagnosis,7,agent-memory
 2605.14906,2026-05-14,recent-2026-Apr-Jul,MemLens: Benchmarking Multimodal Long-Term Memory in Large Vision-Language Models,https://arxiv.org/abs/2605.14906,True,core,update_forgetting,model,retrieval_access;representation_organization;update_forgetting;evaluation_diagnosis;multimodal_embodied_gui;systems_efficiency,benchmark_diagnosis;execution_multimodal,6,agent-memory
 2605.15128,2026-05-14,recent-2026-Apr-Jul,MemEye: A Visual-Centric Evaluation Framework for Multimodal Agent Memory,https://arxiv.org/abs/2605.15128,True,supporting,evaluation,model,retrieval_access;representation_organization;evaluation_diagnosis;multimodal_embodied_gui;systems_efficiency,benchmark_diagnosis;execution_multimodal,3,agent-memory
 2605.15338,2026-05-14,recent-2026-Apr-Jul,Hidden in Memory: Sleeper Memory Poisoning in LLM Agents,https://arxiv.org/abs/2605.15338,True,supporting,security_privacy,model,retrieval_access;evaluation_diagnosis;security_privacy_trust,security_governance,4,long-term-memory
 2605.15701,2026-05-15,recent-2026-Apr-Jul,H-Mem: A Novel Memory Mechanism for Evolving and Retrieving Agent Memory via a Hybrid Structure,https://arxiv.org/abs/2605.15701,True,core,update_forgetting,model,retrieval_access;representation_organization;evaluation_diagnosis;systems_efficiency,retrieval_representation;experience_skill_policy,2,agent-memory;memory-evaluation
 2605.15710,2026-05-15,recent-2026-Apr-Jul,SMMBench: A Benchmark for Source-Distributed Multimodal Agent Memory,https://arxiv.org/abs/2605.15710,True,supporting,evaluation,model,retrieval_access;update_forgetting;evaluation_diagnosis;multimodal_embodied_gui,benchmark_diagnosis;execution_multimodal,4,agent-memory;memory-evaluation
 2605.15759,2026-05-15,recent-2026-Apr-Jul,DimMem: Dimensional Structuring for Efficient Long-Term Agent Memory,https://arxiv.org/abs/2605.15759,True,core,update_forgetting,model,retrieval_access;representation_organization;update_forgetting;evaluation_diagnosis;systems_efficiency,systems_efficiency,5,agent-memory;long-term-memory
 2605.16233,2026-05-15,recent-2026-Apr-Jul,FORGE: Self-Evolving Agent Memory With No Weight Updates via Population Broadcast,https://arxiv.org/abs/2605.16233,True,supporting,evaluation,model,representation_organization;update_forgetting;experience_skill_learning;evaluation_diagnosis;security_privacy_trust;systems_efficiency,update_forgetting;experience_skill_policy,8,agent-memory
 2605.16481,2026-05-15,recent-2026-Apr-Jul,"Visual Agentic Memory: Enabling Online Long Video Understanding via Online Indexing, Hierarchical Memory, and Agentic Retrieval",https://arxiv.org/abs/2605.16481,True,core,update_forgetting,model,retrieval_access;representation_organization;update_forgetting;execution_state;multimodal_embodied_gui,retrieval_representation;execution_multimodal,5,agent-memory
 2605.16746,2026-05-16,recent-2026-Apr-Jul,State Contamination in Memory-Augmented LLM Agents,https://arxiv.org/abs/2605.16746,True,supporting,security_privacy,model,retrieval_access;representation_organization;update_forgetting;execution_state;security_privacy_trust;multi_agent_shared;systems_efficiency,security_governance,6,long-term-memory
 2605.16858,2026-05-16,recent-2026-Apr-Jul,Pedestrian-Aware LLM-Driven Behavioral Planning for Autonomous Vehicles,https://arxiv.org/abs/2605.16858,False,core,execution_state,model,representation_organization;execution_state;experience_skill_learning;evaluation_diagnosis,,7,episodic-memory
 2605.23986,2026-05-16,recent-2026-Apr-Jul,MemForest: An Efficient Agent Memory System with Hierarchical Temporal Indexing,https://arxiv.org/abs/2605.23986,True,core,update_forgetting,model,representation_organization;update_forgetting;evaluation_diagnosis;systems_efficiency,retrieval_representation;systems_efficiency,8,agent-memory;memory-evaluation
 2605.17348,2026-05-17,recent-2026-Apr-Jul,"Taming ""Zombie'' Agents: A Markov State-Aware Framework for Resilient Multi-Agent Evolution",https://arxiv.org/abs/2605.17348,False,core,update_forgetting,model,multi_agent_shared;systems_efficiency,personal_shared,5,agent-memory
 2605.17596,2026-05-17,recent-2026-Apr-Jul,"NeuSymMS: A Hybrid Neuro-Symbolic Memory System for Persistent, Self-Curating LLM Agents",https://arxiv.org/abs/2605.17596,True,core,update_forgetting,model,retrieval_access;representation_organization;security_privacy_trust,,0,long-term-memory
 2605.17625,2026-05-17,recent-2026-Apr-Jul,Episodic-Semantic Memory Architecture for Long-Horizon Scientific Agents,https://arxiv.org/abs/2605.17625,True,core,update_forgetting,model,retrieval_access;update_forgetting;execution_state;evaluation_diagnosis;systems_efficiency,execution_multimodal,7,agent-memory
 2605.17641,2026-05-17,recent-2026-Apr-Jul,Causal Intervention-Based Memory Selection for Long-Horizon LLM Agents,https://arxiv.org/abs/2605.17641,True,core,update_forgetting,model,retrieval_access;representation_organization;update_forgetting;execution_state;experience_skill_learning;evaluation_diagnosis,execution_multimodal,6,long-term-memory
 2605.18284,2026-05-18,recent-2026-Apr-Jul,CommitDistill: A Lightweight Knowledge-Centric Memory Layer for Software Repositories,https://arxiv.org/abs/2605.18284,True,core,execution_state,model,retrieval_access;representation_organization;experience_skill_learning;evaluation_diagnosis;security_privacy_trust;multimodal_embodied_gui;systems_efficiency,experience_skill_policy;systems_efficiency,6,agent-memory
 2605.18421,2026-05-18,recent-2026-Apr-Jul,EvoMemBench: Benchmarking Agent Memory from a Self-Evolving Perspective,https://arxiv.org/abs/2605.18421,True,core,update_forgetting,model,retrieval_access;representation_organization;update_forgetting;execution_state;experience_skill_learning;evaluation_diagnosis,benchmark_diagnosis;experience_skill_policy,6,agent-memory;long-term-memory
 2605.18652,2026-05-18,recent-2026-Apr-Jul,MementoGUI: Learning Agentic Multimodal Memory Control for Long-Horizon GUI Agents,https://arxiv.org/abs/2605.18652,True,core,update_forgetting,model,retrieval_access;execution_state;evaluation_diagnosis;multimodal_embodied_gui;systems_efficiency,experience_skill_policy;execution_multimodal,5,agent-memory
 2605.18930,2026-05-18,recent-2026-Apr-Jul,OEP: Poisoning Self-Evolving LLM Agents via Locally Correct but Non-Transferable Experiences,https://arxiv.org/abs/2605.18930,False,core,update_forgetting,model,representation_organization;update_forgetting;experience_skill_learning;evaluation_diagnosis;security_privacy_trust,experience_skill_policy;security_governance,6,agent-memory
 2605.19952,2026-05-19,recent-2026-Apr-Jul,Rethinking How to Remember: Beyond Atomic Facts in Lifelong LLM Agent Memory,https://arxiv.org/abs/2605.19952,True,core,update_forgetting,model,retrieval_access;representation_organization;systems_efficiency,benchmark_diagnosis,6,agent-memory
 2605.20616,2026-05-20,recent-2026-Apr-Jul,Auto-Dreamer: Learning Offline Memory Consolidation for Language Agents,https://arxiv.org/abs/2605.20616,True,core,update_forgetting,model,retrieval_access;representation_organization;update_forgetting;experience_skill_learning;security_privacy_trust,update_forgetting;experience_skill_policy,4,agent-memory
 2605.20724,2026-05-20,recent-2026-Apr-Jul,CALMem : Application-Layer Dual Memory for Conversational AI,https://arxiv.org/abs/2605.20724,True,core,execution_state,model,retrieval_access;representation_organization;multimodal_embodied_gui;systems_efficiency,,0,episodic-memory
 2605.20833,2026-05-20,recent-2026-Apr-Jul,MemGym: a Long-Horizon Memory Environment for LLM Agents,https://arxiv.org/abs/2605.20833,True,core,execution_state,model,retrieval_access;update_forgetting;execution_state;evaluation_diagnosis;multimodal_embodied_gui;systems_efficiency,execution_multimodal,3,agent-memory;memory-evaluation
 2605.21463,2026-05-20,recent-2026-Apr-Jul,Mem-$π$: Adaptive Memory through Learning When and What to Generate,https://arxiv.org/abs/2605.21463,True,core,execution_state,model,retrieval_access;experience_skill_learning;evaluation_diagnosis;multimodal_embodied_gui,experience_skill_policy,6,episodic-memory
 2605.22411,2026-05-21,recent-2026-Apr-Jul,DeferMem: Query-Time Evidence Distillation via Reinforcement Learning for Long-Term Memory QA,https://arxiv.org/abs/2605.22411,True,core,execution_state,model,retrieval_access;representation_organization;experience_skill_learning;systems_efficiency,experience_skill_policy,2,long-term-memory
 2605.22721,2026-05-21,recent-2026-Apr-Jul,Self-Evolving Multi-Agent Systems via Decentralized Memory,https://arxiv.org/abs/2605.22721,True,core,execution_state,model,update_forgetting;experience_skill_learning;evaluation_diagnosis;security_privacy_trust;multi_agent_shared;multimodal_embodied_gui;systems_efficiency,experience_skill_policy;personal_shared,6,long-term-memory
 2605.23067,2026-05-21,recent-2026-Apr-Jul,What Training Data Teaches RL Memory Agents: An Empirical Study of Curriculum Effects in Memory-Augmented QA,https://arxiv.org/abs/2605.23067,True,core,execution_state,model,experience_skill_learning;evaluation_diagnosis,benchmark_diagnosis;experience_skill_policy,3,agent-memory
 2605.23723,2026-05-22,recent-2026-Apr-Jul,MemAudit: Post-hoc Auditing of Poisoned Agent Memory via Causal Attribution and Structural Anomaly Detection,https://arxiv.org/abs/2605.23723,True,core,execution_state,model,retrieval_access;execution_state;evaluation_diagnosis;security_privacy_trust,security_governance,2,agent-memory;long-term-memory
 2605.24941,2026-05-24,recent-2026-Apr-Jul,Memory-Induced Tool-Drift in LLM Agents,https://arxiv.org/abs/2605.24941,True,core,execution_state,model,evaluation_diagnosis;security_privacy_trust;systems_efficiency,,5,long-term-memory
 2605.25002,2026-05-24,recent-2026-Apr-Jul,MemMark: State-Evolution Attribution Watermarking for Agent Long-Term Memory Systems,https://arxiv.org/abs/2605.25002,True,core,execution_state,model,update_forgetting;evaluation_diagnosis;security_privacy_trust;multimodal_embodied_gui;systems_efficiency,,5,agent-memory
 2605.26154,2026-05-24,recent-2026-Apr-Jul,MemMorph: Tool Hijacking in LLM Agents via Memory Poisoning,https://arxiv.org/abs/2605.26154,True,core,execution_state,model,experience_skill_learning;evaluation_diagnosis;security_privacy_trust;multimodal_embodied_gui,security_governance,6,long-term-memory
 2605.25869,2026-05-25,recent-2026-Apr-Jul,Mitigating Provenance-Role Collapse in Long-Term Agents via Typed Memory Representation,https://arxiv.org/abs/2605.25869,True,core,execution_state,model,retrieval_access;representation_organization;security_privacy_trust;systems_efficiency,retrieval_representation;security_governance,5,long-term-memory
 2605.26252,2026-05-25,recent-2026-Apr-Jul,Is Agent Memory a Database? Rethinking Data Foundations for Long-Term AI Agent Memory,https://arxiv.org/abs/2605.26252,True,core,execution_state,model,retrieval_access;representation_organization;update_forgetting;execution_state;security_privacy_trust;systems_efficiency,benchmark_diagnosis;retrieval_representation,1,agent-memory
 2605.26256,2026-05-25,recent-2026-Apr-Jul,Personalizing Embodied Multimodal Large Language Model Agents over Long-term User Interactions,https://arxiv.org/abs/2605.26256,False,core,update_forgetting,model,retrieval_access;representation_organization;update_forgetting;experience_skill_learning;evaluation_diagnosis;multimodal_embodied_gui,execution_multimodal;personal_shared,2,episodic-memory
 2605.27366,2026-05-26,recent-2026-Apr-Jul,"MUSE-Autoskill: Self-Evolving Agents via Skill Creation, Memory, Management, and Evaluation",https://arxiv.org/abs/2605.27366,True,core,update_forgetting,model,retrieval_access;experience_skill_learning;evaluation_diagnosis,benchmark_diagnosis;experience_skill_policy,6,agent-memory
 2605.27760,2026-05-26,recent-2026-Apr-Jul,SkillGrad: Optimizing Agent Skills Like Gradient Descent,https://arxiv.org/abs/2605.27760,False,core,update_forgetting,model,representation_organization;update_forgetting;execution_state;experience_skill_learning;evaluation_diagnosis,experience_skill_policy,5,long-term-memory
 2605.27762,2026-05-26,recent-2026-Apr-Jul,PEAM: Parametric Embodied Agent Memory through Contrastive Internalization of Experience in Minecraft,https://arxiv.org/abs/2605.27762,True,core,update_forgetting,model,retrieval_access;update_forgetting;execution_state;experience_skill_learning;evaluation_diagnosis;security_privacy_trust;multimodal_embodied_gui,experience_skill_policy;execution_multimodal,5,agent-memory
 2605.27825,2026-05-27,recent-2026-Apr-Jul,MRMMIA: Membership Inference Attacks on Memory in Chat Agents,https://arxiv.org/abs/2605.27825,True,core,security_privacy,model,retrieval_access;representation_organization;evaluation_diagnosis;security_privacy_trust,security_governance,4,agent-memory
 2605.28046,2026-05-27,recent-2026-Apr-Jul,MemCog: From Memory-as-Tool to Memory-as-Cognition in Conversational Agents,https://arxiv.org/abs/2605.28046,True,core,execution_state,model,retrieval_access;evaluation_diagnosis;multimodal_embodied_gui,,4,agent-memory
 2605.29341,2026-05-28,recent-2026-Apr-Jul,WorldMemArena: Evaluating Multimodal Agent Memory Through Action-World Interaction,https://arxiv.org/abs/2605.29341,True,core,update_forgetting,model,retrieval_access;update_forgetting;execution_state;evaluation_diagnosis;multimodal_embodied_gui;systems_efficiency,benchmark_diagnosis;execution_multimodal,5,agent-memory
 2605.29630,2026-05-28,recent-2026-Apr-Jul,Entity-Collision: A Stratified Protocol for Attributing Retrieval Lift in Agent Memory,https://arxiv.org/abs/2605.29630,True,core,update_forgetting,model,retrieval_access;evaluation_diagnosis;security_privacy_trust;systems_efficiency,retrieval_representation,7,agent-memory;memory-evaluation
 2605.29640,2026-05-28,recent-2026-Apr-Jul,VikingMem: A Memory Base Management System for Stateful LLM-based Applications,https://arxiv.org/abs/2605.29640,True,core,update_forgetting,model,retrieval_access;update_forgetting;evaluation_diagnosis;systems_efficiency,,6,agent-memory;memory-evaluation
 2605.29960,2026-05-28,recent-2026-Apr-Jul,Hijacking Agent Memory: Stealthy Trojan Attacks Through Conversational Interaction,https://arxiv.org/abs/2605.29960,True,core,update_forgetting,model,evaluation_diagnosis;security_privacy_trust,security_governance,5,agent-memory;long-term-memory
 2605.30058,2026-05-28,recent-2026-Apr-Jul,HEART-Bench: Do LLM Agents Exhibit Human-like Psychology?,https://arxiv.org/abs/2605.30058,False,core,experience_learning,model,representation_organization;update_forgetting;execution_state;evaluation_diagnosis;multimodal_embodied_gui,,2,episodic-memory
 2605.30621,2026-05-28,recent-2026-Apr-Jul,Harness Updating Is Not Harness Benefit: Disentangling Evolution Capabilities in Self-Evolving LLM Agents,https://arxiv.org/abs/2605.30621,False,core,update_forgetting,model,update_forgetting;execution_state;experience_skill_learning;systems_efficiency,experience_skill_policy,2,procedural-memory
 2607.00017,2026-05-28,recent-2026-Apr-Jul,Learning User-Aware Recall: Personalized Retrieval in Long-Term Conversational Memory,https://arxiv.org/abs/2607.00017,True,core,experience_learning,model,retrieval_access;evaluation_diagnosis;multimodal_embodied_gui,retrieval_representation;experience_skill_policy;personal_shared,5,long-term-memory
 2605.30690,2026-05-29,recent-2026-Apr-Jul,ElasticMem: Latent Memory as a Learnable Resource for LLM Agents,https://arxiv.org/abs/2605.30690,True,core,representation,model,retrieval_access;experience_skill_learning;evaluation_diagnosis;multimodal_embodied_gui;systems_efficiency,experience_skill_policy,8,long-term-memory
 2605.30711,2026-05-29,recent-2026-Apr-Jul,SAGE: A Novelty Gate for Efficient Memory Evolution in Agentic LLMs,https://arxiv.org/abs/2605.30711,True,core,representation,model,retrieval_access;systems_efficiency,systems_efficiency,3,agent-memory
 2605.30858,2026-05-29,recent-2026-Apr-Jul,ForecastCompass: Guiding Agentic Forecasting with Adaptive Factor Memory,https://arxiv.org/abs/2605.30858,True,core,representation,model,retrieval_access;representation_organization;experience_skill_learning;multimodal_embodied_gui,experience_skill_policy;execution_multimodal,2,agent-memory
 2605.31075,2026-05-29,recent-2026-Apr-Jul,Task-Focused Memorization for Multimodal Agents,https://arxiv.org/abs/2605.31075,True,core,representation,model,experience_skill_learning;evaluation_diagnosis;multimodal_embodied_gui,execution_multimodal,4,agent-memory
 2606.00619,2026-05-30,recent-2026-Apr-Jul,MemPro: Agentic Memory Systems as Evolvable Programs,https://arxiv.org/abs/2606.00619,True,core,representation,model,retrieval_access;representation_organization;execution_state;evaluation_diagnosis;multimodal_embodied_gui;systems_efficiency,experience_skill_policy,6,agent-memory
 2606.00734,2026-05-30,recent-2026-Apr-Jul,EMA: Approximate Nearest Neighbor Search with General Attribute Filtering and Dynamic Updates,https://arxiv.org/abs/2606.00734,False,core,representation,model,retrieval_access;representation_organization;update_forgetting;evaluation_diagnosis;multimodal_embodied_gui;systems_efficiency,retrieval_representation;update_forgetting,4,agent-memory
 2606.00756,2026-05-30,recent-2026-Apr-Jul,CoMIC: Collaborative Memory and Insights Circulation for Long-Horizon LLM Agents in Cloud-Edge Systems,https://arxiv.org/abs/2606.00756,True,core,representation,model,representation_organization;update_forgetting;execution_state;experience_skill_learning;evaluation_diagnosis;multi_agent_shared;multimodal_embodied_gui;systems_efficiency,execution_multimodal;systems_efficiency;personal_shared,2,long-term-memory
 2606.01138,2026-05-31,recent-2026-Apr-Jul,memorywire: A Vendor-Neutral Wire Format for Agent Memory Operations,https://arxiv.org/abs/2606.01138,True,core,representation,model,retrieval_access;update_forgetting;experience_skill_learning;evaluation_diagnosis;security_privacy_trust;systems_efficiency,systems_efficiency,6,agent-memory
 2606.01528,2026-06-01,recent-2026-Apr-Jul,Joint Agent Memory and Exploration Learning via Novelty Signals,https://arxiv.org/abs/2606.01528,True,core,representation,model,evaluation_diagnosis;multimodal_embodied_gui;systems_efficiency,experience_skill_policy,5,agent-memory
 2606.01722,2026-06-01,recent-2026-Apr-Jul,Post-Deterministic Distributed Systems: A New Foundation for Trustworthy Autonomous Infrastructure,https://arxiv.org/abs/2606.01722,False,core,representation,model,retrieval_access;representation_organization;security_privacy_trust;multimodal_embodied_gui,retrieval_representation;security_governance,1,agent-memory
 2606.02812,2026-06-01,recent-2026-Apr-Jul,Traj-Evolve: A Self-Evolving Multi-Agent System for Patient Trajectory Modeling in Lung Cancer Early Detection,https://arxiv.org/abs/2606.02812,False,core,representation,model,retrieval_access;representation_organization;execution_state;experience_skill_learning;evaluation_diagnosis;multi_agent_shared;multimodal_embodied_gui,experience_skill_policy;personal_shared,3,agent-memory
 2606.03197,2026-06-02,recent-2026-Apr-Jul,MemTrain: Self-Supervised Context Memory Training,https://arxiv.org/abs/2606.03197,True,core,update_forgetting,model,retrieval_access;update_forgetting;execution_state;experience_skill_learning;evaluation_diagnosis;systems_efficiency,,4,agent-memory
 2606.03329,2026-06-02,recent-2026-Apr-Jul,InfoMem: Training Long-Context Memory Agents with Answer-Conditioned Information Gain,https://arxiv.org/abs/2606.03329,True,core,update_forgetting,model,retrieval_access;experience_skill_learning;evaluation_diagnosis;systems_efficiency,,5,agent-memory
 2606.03374,2026-06-02,recent-2026-Apr-Jul,eMEM: A Hybrid Spatio-Temporal Memory System For Embodied Agents,https://arxiv.org/abs/2606.03374,True,core,execution_state,model,retrieval_access;representation_organization;update_forgetting;execution_state;evaluation_diagnosis;multimodal_embodied_gui;systems_efficiency,execution_multimodal,6,agent-memory;memory-evaluation
 2606.04120,2026-06-02,recent-2026-Apr-Jul,SaliMory: Orchestrating Cognitive Memory for Conversational Agents,https://arxiv.org/abs/2606.04120,True,core,update_forgetting,model,retrieval_access;representation_organization;update_forgetting;experience_skill_learning,,5,agent-memory
 2606.04315,2026-06-03,recent-2026-Apr-Jul,Exploring Cross-Scenario Generality of Agentic Memory Systems: Diagnostics and a Strong Baseline,https://arxiv.org/abs/2606.04315,True,core,execution_state,model,retrieval_access;execution_state;evaluation_diagnosis;systems_efficiency,benchmark_diagnosis,2,agent-memory
 2606.04329,2026-06-03,recent-2026-Apr-Jul,From Untrusted Input to Trusted Memory: A Systematic Study of Memory Poisoning Attacks in LLM Agents,https://arxiv.org/abs/2606.04329,True,core,security_privacy,model,retrieval_access;evaluation_diagnosis;security_privacy_trust,benchmark_diagnosis;security_governance,3,long-term-memory
 2606.04555,2026-06-03,recent-2026-Apr-Jul,Temporal Order Matters for Agentic Memory: Segment Trees for Long-Horizon Agents,https://arxiv.org/abs/2606.04555,True,core,update_forgetting,model,retrieval_access;representation_organization;update_forgetting;execution_state;evaluation_diagnosis;systems_efficiency,execution_multimodal,4,agent-memory;memory-evaluation
 2606.04628,2026-06-03,recent-2026-Apr-Jul,RAMPART: Registry-based Agentic Memory with Priority-Aware Runtime Transformation,https://arxiv.org/abs/2606.04628,True,core,update_forgetting,model,retrieval_access;representation_organization;evaluation_diagnosis;security_privacy_trust;systems_efficiency,systems_efficiency,6,agent-memory
 2606.04780,2026-06-03,recent-2026-Apr-Jul,PersonaTree: Structured Lifecycle Memory for Person Understanding in LLM Agents,https://arxiv.org/abs/2606.04780,True,core,experience_learning,model,retrieval_access;representation_organization;update_forgetting;evaluation_diagnosis;multimodal_embodied_gui;systems_efficiency,retrieval_representation;update_forgetting,3,agent-memory;long-term-memory;memory-evaluation
 2606.05513,2026-06-03,recent-2026-Apr-Jul,EpiEvolve: Self-Evolving Agents for Streaming Pandemic Forecasting under Regime Shifts,https://arxiv.org/abs/2606.05513,False,core,update_forgetting,model,retrieval_access;representation_organization;experience_skill_learning;evaluation_diagnosis;security_privacy_trust,experience_skill_policy,5,episodic-memory
 2606.28349,2026-06-03,recent-2026-Apr-Jul,HMARS: A Hierarchical Multi-Agent Memory System for Long-Context Reasoning,https://arxiv.org/abs/2606.28349,True,core,representation,model,retrieval_access;representation_organization;evaluation_diagnosis;multi_agent_shared,retrieval_representation;personal_shared,4,agent-memory
 2606.05646,2026-06-04,recent-2026-Apr-Jul,Enhancing Software Engineering Through Closed-Loop Memory Optimization,https://arxiv.org/abs/2606.05646,True,core,update_forgetting,model,experience_skill_learning;evaluation_diagnosis;systems_efficiency,,4,episodic-memory
 2606.05684,2026-06-04,recent-2026-Apr-Jul,AdaMEM: Test-Time Adaptive Memory for Language Agents,https://arxiv.org/abs/2606.05684,True,core,update_forgetting,model,retrieval_access;execution_state;experience_skill_learning;multimodal_embodied_gui;systems_efficiency,experience_skill_policy,6,agent-memory
 2606.06054,2026-06-04,recent-2026-Apr-Jul,Beyond Similarity: Trustworthy Memory Search for Personal AI Agents,https://arxiv.org/abs/2606.06054,True,core,update_forgetting,model,retrieval_access;representation_organization;update_forgetting;evaluation_diagnosis;security_privacy_trust;systems_efficiency,retrieval_representation;security_governance,2,agent-memory
 2606.06090,2026-06-04,recent-2026-Apr-Jul,Beyond Semantic Organization: Memory as Execution State Management for Long-Horizon Agents,https://arxiv.org/abs/2606.06090,True,core,execution_state,model,retrieval_access;representation_organization;execution_state;security_privacy_trust;multimodal_embodied_gui;systems_efficiency,retrieval_representation;execution_multimodal,6,agent-memory
 2606.06240,2026-06-04,recent-2026-Apr-Jul,TOKI: A Bitemporal Operator Algebra for Contradiction Resolution in LLM-Agent Persistent Memory,https://arxiv.org/abs/2606.06240,True,core,update_forgetting,model,update_forgetting;security_privacy_trust,,2,long-term-memory
 2606.06448,2026-06-04,recent-2026-Apr-Jul,Agent Memory: Characterization and System Implications of Stateful Long-Horizon Workloads,https://arxiv.org/abs/2606.06448,True,core,execution_state,model,retrieval_access;update_forgetting;execution_state;evaluation_diagnosis;systems_efficiency,execution_multimodal,2,agent-memory
 2606.07402,2026-06-05,recent-2026-Apr-Jul,M$^3$Exam: Benchmarking Multimodal Memory for Realistic User-Agent Interactions,https://arxiv.org/abs/2606.07402,True,core,access,model,retrieval_access;representation_organization;evaluation_diagnosis;multimodal_embodied_gui;systems_efficiency,benchmark_diagnosis;execution_multimodal,4,memory-evaluation
 2606.09900,2026-06-05,recent-2026-Apr-Jul,"Less Context, More Accuracy: A Bi-Temporal Memory Engine for LLM Agents Where a Lean Retrieved Context Beats the Full History",https://arxiv.org/abs/2606.09900,True,core,update_forgetting,model,retrieval_access;representation_organization;update_forgetting;evaluation_diagnosis;security_privacy_trust;systems_efficiency,retrieval_representation,7,long-term-memory;memory-evaluation
 2606.24896,2026-06-05,recent-2026-Apr-Jul,Why Memory Components Fail: Eight Years of License and Sustainability Events in Open-Source Data Infrastructure,https://arxiv.org/abs/2606.24896,True,core,access,model,representation_organization;experience_skill_learning;security_privacy_trust,retrieval_representation,1,agent-memory
 2606.08367,2026-06-06,recent-2026-Apr-Jul,Emergence World: A Platform for Evaluating Long-Horizon Multi-Agent Autonomy,https://arxiv.org/abs/2606.08367,False,core,execution_state,model,retrieval_access;execution_state;evaluation_diagnosis;security_privacy_trust;multi_agent_shared,benchmark_diagnosis;execution_multimodal;personal_shared,3,long-term-memory
 2606.09198,2026-06-08,recent-2026-Apr-Jul,MASS: Deep Research for Social Sciences with Memory-Augmented Social Simulation,https://arxiv.org/abs/2606.09198,True,core,experience_learning,model,retrieval_access;representation_organization;update_forgetting;execution_state;multimodal_embodied_gui,retrieval_representation,4,agent-memory
 2606.09461,2026-06-08,recent-2026-Apr-Jul,H2HMem: A Multimodal Memory Benchmark for Agents in Human-Human Interactions,https://arxiv.org/abs/2606.09461,True,core,representation,model,retrieval_access;update_forgetting;evaluation_diagnosis;multimodal_embodied_gui,benchmark_diagnosis;execution_multimodal,3,memory-evaluation
 2606.09483,2026-06-08,recent-2026-Apr-Jul,Memory Beyond Recall: A Dual-Process Cognitive Memory System for Self-Evolving LLM Agents,https://arxiv.org/abs/2606.09483,True,core,update_forgetting,model,retrieval_access;representation_organization;update_forgetting;experience_skill_learning;evaluation_diagnosis,retrieval_representation;experience_skill_policy,2,agent-memory;long-term-memory
 2606.09774,2026-06-08,recent-2026-Apr-Jul,Auto-Configuring Scientific Simulators with Lightweight Coding-Agent Adapters,https://arxiv.org/abs/2606.09774,False,core,execution_state,model,retrieval_access;experience_skill_learning;systems_efficiency,experience_skill_policy,2,agent-memory
 2606.10062,2026-06-08,recent-2026-Apr-Jul,Deployment-Time Memorization in Foundation-Model Agents,https://arxiv.org/abs/2606.10062,True,supporting,evaluation,model,retrieval_access;update_forgetting;evaluation_diagnosis;security_privacy_trust;systems_efficiency,,5,agent-memory
 2606.10299,2026-06-09,recent-2026-Apr-Jul,What Spatial Memory Must Store: Occlusion as the Test for Language-Agent Memory,https://arxiv.org/abs/2606.10299,True,core,execution_state,model,retrieval_access,execution_multimodal,7,agent-memory
 2606.10423,2026-06-09,recent-2026-Apr-Jul,WebChallenger: A Reliable and Efficient Generalist Web Agent,https://arxiv.org/abs/2606.10423,False,core,execution_state,model,representation_organization;experience_skill_learning;multimodal_embodied_gui;systems_efficiency,execution_multimodal;systems_efficiency,4,long-term-memory
 2606.10532,2026-06-09,recent-2026-Apr-Jul,ActiveMem: Distributed Active Memory for Long-Horizon LLM Reasoning,https://arxiv.org/abs/2606.10532,True,core,execution_state,model,retrieval_access;update_forgetting;execution_state,execution_multimodal,3,agent-memory
 2606.10577,2026-06-09,recent-2026-Apr-Jul,AgenticNav: Zero-Shot Vision-and-Language Navigation as a Tool-Calling Harness,https://arxiv.org/abs/2606.10577,False,core,execution_state,model,retrieval_access;execution_state;experience_skill_learning;evaluation_diagnosis;multimodal_embodied_gui,execution_multimodal,6,agent-memory
 2606.10677,2026-06-09,recent-2026-Apr-Jul,Infini Memory: Maintainable Topic Documents for Long-Term LLM Agent Memory,https://arxiv.org/abs/2606.10677,True,core,execution_state,model,retrieval_access;representation_organization;update_forgetting;evaluation_diagnosis;systems_efficiency,retrieval_representation,3,agent-memory;long-term-memory
 2606.11680,2026-06-10,recent-2026-Apr-Jul,Organize then Retrieve: Hierarchical Memory Navigation for Efficient Agents,https://arxiv.org/abs/2606.11680,True,core,execution_state,model,retrieval_access;representation_organization;execution_state;experience_skill_learning;multimodal_embodied_gui;systems_efficiency,retrieval_representation;execution_multimodal;systems_efficiency,5,agent-memory
 2606.12290,2026-06-10,recent-2026-Apr-Jul,Selection Integrity for LLM Graph Memory: An Accumulability Criterion for Information-Flow-Blind Retrieval,https://arxiv.org/abs/2606.12290,True,core,execution_state,model,retrieval_access;representation_organization;security_privacy_trust;systems_efficiency,retrieval_representation;security_governance,0,agent-memory
 2606.12703,2026-06-10,recent-2026-Apr-Jul,SMSR: Certified Defence Against Runtime Memory Poisoning in Persistent LLM Agent Systems,https://arxiv.org/abs/2606.12703,True,core,execution_state,model,retrieval_access;evaluation_diagnosis;security_privacy_trust,security_governance;systems_efficiency,3,long-term-memory
 2606.12852,2026-06-11,recent-2026-Apr-Jul,WISE: A Long-Horizon Agent in Minecraft with Why-Which Reasoning,https://arxiv.org/abs/2606.12852,False,core,execution_state,model,retrieval_access;representation_organization;execution_state;multimodal_embodied_gui,execution_multimodal,3,episodic-memory
 2606.12945,2026-06-11,recent-2026-Apr-Jul,Learning What to Remember: A Cognitively Grounded Multi-Factor Value Model for Agentic Memory,https://arxiv.org/abs/2606.12945,True,core,execution_state,model,retrieval_access;update_forgetting;evaluation_diagnosis;systems_efficiency,experience_skill_policy,9,agent-memory
 2606.13177,2026-06-11,recent-2026-Apr-Jul,MemRefine: LLM-Guided Compression for Long-Term Agent Memory,https://arxiv.org/abs/2606.13177,True,core,update_forgetting,model,retrieval_access;experience_skill_learning;evaluation_diagnosis;multimodal_embodied_gui;systems_efficiency,execution_multimodal;systems_efficiency,4,agent-memory
 2606.13392,2026-06-11,recent-2026-Apr-Jul,MiniMax Sparse Attention,https://arxiv.org/abs/2606.13392,False,supporting,execution_state,model,retrieval_access;representation_organization;multimodal_embodied_gui;systems_efficiency,,4,long-term-memory
 2606.14470,2026-06-12,recent-2026-Apr-Jul,"GitOfThoughts: Version-Controlled Reasoning and Agent Memory You Can Replay, Diff, and Merge",https://arxiv.org/abs/2606.14470,True,supporting,execution_state,model,retrieval_access;representation_organization;evaluation_diagnosis;systems_efficiency,,6,agent-memory
 2606.14571,2026-06-12,recent-2026-Apr-Jul,StreamMemBench: Streaming Evaluation of Agent Memory for Future-Oriented Assistance,https://arxiv.org/abs/2606.14571,True,supporting,execution_state,model,retrieval_access;execution_state;experience_skill_learning;evaluation_diagnosis,benchmark_diagnosis,4,agent-memory;memory-evaluation
 2606.15071,2026-06-13,recent-2026-Apr-Jul,Quantum learning with a single-atom sensor,https://arxiv.org/abs/2606.15071,False,supporting,execution_state,model,systems_efficiency,experience_skill_policy,0,agent-memory
 2606.15609,2026-06-14,recent-2026-Apr-Jul,FragFuse: Bypassing Access Control of Large Language Model Agents via Memory-Based Query Fragmentation and Fusion,https://arxiv.org/abs/2606.15609,True,supporting,execution_state,model,retrieval_access;evaluation_diagnosis;security_privacy_trust,,6,agent-memory;long-term-memory
 2606.15903,2026-06-14,recent-2026-Apr-Jul,Control-Plane Placement Shapes Forgetting: An Architectural Study of Agent Memory Across Thirteen System Configurations,https://arxiv.org/abs/2606.15903,True,supporting,execution_state,model,retrieval_access;update_forgetting;evaluation_diagnosis;security_privacy_trust;systems_efficiency,benchmark_diagnosis;update_forgetting,8,agent-memory
 2606.16903,2026-06-15,recent-2026-Apr-Jul,Directory-Aware Query and Maintenance in Vector Databases,https://arxiv.org/abs/2606.16903,False,supporting,execution_state,model,retrieval_access;representation_organization;update_forgetting;evaluation_diagnosis;systems_efficiency,retrieval_representation,4,agent-memory
 2606.17328,2026-06-15,recent-2026-Apr-Jul,MemTrace: Probing What Final Accuracy Misses in Long-Term Memory,https://arxiv.org/abs/2606.17328,True,supporting,execution_state,model,retrieval_access;execution_state;evaluation_diagnosis;systems_efficiency,,4,long-term-memory
 2606.17628,2026-06-16,recent-2026-Apr-Jul,OPD-Evolver: Cultivating Holistic Agent Evolver via On-Policy Distillation,https://arxiv.org/abs/2606.17628,False,supporting,execution_state,model,retrieval_access;representation_organization;experience_skill_learning;evaluation_diagnosis,experience_skill_policy,4,agent-memory
 2606.18356,2026-06-16,recent-2026-Apr-Jul,"SafeClawBench: Separating Semantic, Audit-Evidence, and Sandbox Harm in Tool-Using LLM Agents",https://arxiv.org/abs/2606.18356,False,supporting,execution_state,model,representation_organization;evaluation_diagnosis;security_privacy_trust;multimodal_embodied_gui,,7,long-term-memory
 2606.18406,2026-06-16,recent-2026-Apr-Jul,CoreMem: Riemannian Retrieval and Fisher-Guided Distillation for Long-Term Memory in Dialogue Agents,https://arxiv.org/abs/2606.18406,True,supporting,execution_state,model,retrieval_access;representation_organization;evaluation_diagnosis;multimodal_embodied_gui;systems_efficiency,retrieval_representation;experience_skill_policy;execution_multimodal,5,agent-memory
 2606.18746,2026-06-17,recent-2026-Apr-Jul,What Must Generalist Agents Remember?,https://arxiv.org/abs/2606.18746,True,core,experience_learning,model,execution_state,,0,agent-memory
 2606.18829,2026-06-17,recent-2026-Apr-Jul,GateMem: Benchmarking Memory Governance in Multi-Principal Shared-Memory Agents,https://arxiv.org/abs/2606.18829,True,supporting,shared_memory,model,retrieval_access;representation_organization;update_forgetting;execution_state;evaluation_diagnosis;security_privacy_trust;systems_efficiency,benchmark_diagnosis;security_governance,5,agent-memory;memory-evaluation
 2606.18950,2026-06-17,recent-2026-Apr-Jul,RTSGameBench: An RTS Benchmark for Strategic Reasoning by Vision-Language Models,https://arxiv.org/abs/2606.18950,False,supporting,update_forgetting,model,representation_organization;execution_state;experience_skill_learning;evaluation_diagnosis;multi_agent_shared,benchmark_diagnosis,3,agent-memory
 2606.19409,2026-06-17,recent-2026-Apr-Jul,OpenRath: Session-Centered Runtime State for Agent Systems,https://arxiv.org/abs/2606.19409,False,core,representation,model,evaluation_diagnosis;security_privacy_trust;multi_agent_shared;systems_efficiency,systems_efficiency,3,agent-memory
 2606.19847,2026-06-18,recent-2026-Apr-Jul,AtomMem: Building Simple and Effective Memory System for LLM Agents via Atomic Facts,https://arxiv.org/abs/2606.19847,True,core,representation,model,retrieval_access;representation_organization;update_forgetting;evaluation_diagnosis;systems_efficiency,,2,long-term-memory
 2606.19857,2026-06-18,recent-2026-Apr-Jul,Large Language Models Do Not Always Need Readable Language,https://arxiv.org/abs/2606.19857,False,supporting,representation,model,representation_organization;evaluation_diagnosis;multi_agent_shared;systems_efficiency,,4,agent-memory
 2606.20047,2026-06-18,recent-2026-Apr-Jul,PACMS: Submodular Context Selection as a Pluggable Engine for LLM Agents,https://arxiv.org/abs/2606.20047,False,supporting,update_forgetting,model,retrieval_access;systems_efficiency,,2,long-term-memory
 2606.20515,2026-06-18,recent-2026-Apr-Jul,S-Agent: Spatial Tool-Use Elicits Reasoning for Spatial Intelligence,https://arxiv.org/abs/2606.20515,False,core,representation,model,representation_organization;evaluation_diagnosis;multimodal_embodied_gui,execution_multimodal,6,agent-memory
 2606.20954,2026-06-18,recent-2026-Apr-Jul,Learning What Not to Forget: Long-Horizon Agent Memory from a Few Kilobytes of Learning,https://arxiv.org/abs/2606.20954,True,core,representation,model,retrieval_access;update_forgetting;execution_state;evaluation_diagnosis;systems_efficiency,update_forgetting;experience_skill_policy;execution_multimodal,8,agent-memory
 2606.21144,2026-06-19,recent-2026-Apr-Jul,AdaMem: Learning What to Remember for Personalized Long-Horizon LLM Agents,https://arxiv.org/abs/2606.21144,True,core,representation,model,representation_organization;update_forgetting;execution_state;experience_skill_learning;evaluation_diagnosis;systems_efficiency,experience_skill_policy;execution_multimodal;personal_shared,5,long-term-memory
 2606.21562,2026-06-19,recent-2026-Apr-Jul,Compressing Observation History into Agent Memory: Distilling Transformers into Recurrent Transformers,https://arxiv.org/abs/2606.21562,True,supporting,representation,model,representation_organization;execution_state;multimodal_embodied_gui;systems_efficiency,experience_skill_policy;systems_efficiency,2,agent-memory
 2606.21649,2026-06-19,recent-2026-Apr-Jul,EvoEmbedding: Evolvable Representations for Long-Context Retrieval and Agentic Memory,https://arxiv.org/abs/2606.21649,True,supporting,representation,model,retrieval_access;representation_organization;update_forgetting;evaluation_diagnosis,retrieval_representation;experience_skill_policy,6,agent-memory
 2606.22030,2026-06-20,recent-2026-Apr-Jul,Nous: A Predictive World Model for Long-Term Agent Memory,https://arxiv.org/abs/2606.22030,True,core,experience_learning,model,representation_organization;update_forgetting;evaluation_diagnosis;systems_efficiency,,5,agent-memory;memory-evaluation
 2606.22263,2026-06-20,recent-2026-Apr-Jul,Revelio: Cost-Efficient Agentic Memory Safety Vulnerability Detection For Repository-Scale Codebases,https://arxiv.org/abs/2606.22263,True,supporting,security_privacy,model,evaluation_diagnosis;security_privacy_trust;systems_efficiency,security_governance;systems_efficiency,5,agent-memory
 2606.22844,2026-06-22,recent-2026-Apr-Jul,RaMem: Contextual Reinstatement for Long-term Agentic Memory,https://arxiv.org/abs/2606.22844,True,core,execution_state,model,retrieval_access;representation_organization;experience_skill_learning;evaluation_diagnosis,,7,agent-memory;long-term-memory;memory-evaluation
 2606.22877,2026-06-22,recent-2026-Apr-Jul,DynamicMem: A Long-Horizon Memory Benchmark in Real-World Settings,https://arxiv.org/abs/2606.22877,True,supporting,security_privacy,model,retrieval_access;update_forgetting;execution_state;evaluation_diagnosis;security_privacy_trust;systems_efficiency,benchmark_diagnosis;execution_multimodal,6,memory-evaluation
 2606.23127,2026-06-22,recent-2026-Apr-Jul,"Managing Procedural Memory in LLM Agents: Control, Adaptation, and Evaluation",https://arxiv.org/abs/2606.23127,True,core,evaluation,model,experience_skill_learning;evaluation_diagnosis;multimodal_embodied_gui,benchmark_diagnosis;experience_skill_policy,7,procedural-memory
 2606.23195,2026-06-22,recent-2026-Apr-Jul,Memory Contagion: Cross-Temporal Propagation of Evaluator Bias via Agent Memory,https://arxiv.org/abs/2606.23195,True,supporting,security_privacy,model,retrieval_access;update_forgetting;experience_skill_learning;evaluation_diagnosis;multimodal_embodied_gui,benchmark_diagnosis,4,agent-memory
 2606.23283,2026-06-22,recent-2026-Apr-Jul,Towards Root Memories: Benchmarking and Enhancing Implicit Logical Memory Retrieval for Personalized LLMs,https://arxiv.org/abs/2606.23283,True,core,execution_state,model,retrieval_access;representation_organization;evaluation_diagnosis;systems_efficiency,benchmark_diagnosis;retrieval_representation;personal_shared,5,agent-memory
 2606.24322,2026-06-23,recent-2026-Apr-Jul,"Securing LLM-Agent Long-Term Memory Against Poisoning: Non-Malleable, Origin-Bound Authority with Machine-Checked Guarantees",https://arxiv.org/abs/2606.24322,True,supporting,security_privacy,model,retrieval_access;evaluation_diagnosis;security_privacy_trust,security_governance,6,agent-memory;long-term-memory
 2606.24535,2026-06-23,recent-2026-Apr-Jul,Governed Shared Memory for Multi-Agent LLM Systems,https://arxiv.org/abs/2606.24535,True,core,shared_memory,model,retrieval_access;update_forgetting;evaluation_diagnosis;security_privacy_trust;multi_agent_shared;systems_efficiency,security_governance;personal_shared,7,agent-memory
 2606.24595,2026-06-23,recent-2026-Apr-Jul,MEMPROBE: Probing Long-Term Agent Memory via Hidden User-State Recovery,https://arxiv.org/abs/2606.24595,True,supporting,security_privacy,model,retrieval_access;representation_organization;execution_state;evaluation_diagnosis;security_privacy_trust;systems_efficiency,,5,agent-memory;long-term-memory
 2606.24775,2026-06-23,recent-2026-Apr-Jul,Are We Ready For An Agent-Native Memory System?,https://arxiv.org/abs/2606.24775,True,core,update_forgetting,model,retrieval_access;representation_organization;update_forgetting;execution_state;evaluation_diagnosis;security_privacy_trust;systems_efficiency,,6,agent-memory
 2606.25115,2026-06-23,recent-2026-Apr-Jul,Forget to Improve: On-Device LLM-Agent Continual Learning via Budget-Curated Memory,https://arxiv.org/abs/2606.25115,True,supporting,security_privacy,model,retrieval_access;update_forgetting;experience_skill_learning;evaluation_diagnosis;security_privacy_trust;multimodal_embodied_gui;systems_efficiency,update_forgetting;experience_skill_policy;systems_efficiency,2,agent-memory
 2606.25161,2026-06-23,recent-2026-Apr-Jul,TRUSTMEM: Learning Trustworthy Memory Consolidation for LLM Agents with Long-Term Memory,https://arxiv.org/abs/2606.25161,True,core,update_forgetting,model,update_forgetting;experience_skill_learning;evaluation_diagnosis;security_privacy_trust;multimodal_embodied_gui,update_forgetting;experience_skill_policy;security_governance,7,agent-memory;long-term-memory
 2606.25206,2026-06-23,recent-2026-Apr-Jul,RAVEN: Long-Horizon Reasoning & Navigation with a Visuo-Spatio-Temporal Memory,https://arxiv.org/abs/2606.25206,True,core,execution_state,model,retrieval_access;representation_organization;execution_state;evaluation_diagnosis;multimodal_embodied_gui;systems_efficiency,execution_multimodal,2,agent-memory
 2606.26627,2026-06-25,recent-2026-Apr-Jul,Agents That Know Too Much: A Data-Centric Survey of Privacy in LLM Agents,https://arxiv.org/abs/2606.26627,False,supporting,security_privacy,model,retrieval_access;representation_organization;evaluation_diagnosis;security_privacy_trust,benchmark_diagnosis;security_governance,2,agent-memory
 2606.26790,2026-06-25,recent-2026-Apr-Jul,OPID: On-Policy Skill Distillation for Agentic Reinforcement Learning,https://arxiv.org/abs/2606.26790,False,core,update_forgetting,model,retrieval_access;representation_organization;execution_state;experience_skill_learning;multimodal_embodied_gui;systems_efficiency,experience_skill_policy,6,procedural-memory
 2606.27499,2026-06-25,recent-2026-Apr-Jul,DMV-Bench: Diagnosing Long-Horizon Multimodal Agents' Visual Memory with Incidental Cue Injection,https://arxiv.org/abs/2606.27499,True,core,execution_state,model,retrieval_access;execution_state;evaluation_diagnosis;security_privacy_trust;multimodal_embodied_gui,benchmark_diagnosis;execution_multimodal,6,agent-memory
 2606.28781,2026-06-27,recent-2026-Apr-Jul,HyphaeDB: A Living Knowledge Topology for Agent-First Memory,https://arxiv.org/abs/2606.28781,True,core,execution_state,model,retrieval_access;representation_organization;multi_agent_shared;systems_efficiency,systems_efficiency,0,agent-memory
 2606.28971,2026-06-27,recent-2026-Apr-Jul,Self-Evolving Agentic Image Restoration via Deliberate Planning and Intuitive Execution,https://arxiv.org/abs/2606.28971,False,core,update_forgetting,model,retrieval_access;representation_organization;execution_state;experience_skill_learning;evaluation_diagnosis;multimodal_embodied_gui;systems_efficiency,experience_skill_policy,3,episodic-memory
 2606.29774,2026-06-29,recent-2026-Apr-Jul,Analytic Concept-Centric Memory for Agentic Embodied Manipulation,https://arxiv.org/abs/2606.29774,True,core,update_forgetting,model,retrieval_access;representation_organization;execution_state;experience_skill_learning;evaluation_diagnosis;multimodal_embodied_gui,execution_multimodal,5,agent-memory;memory-evaluation
 2606.29778,2026-06-29,recent-2026-Apr-Jul,Mandol: An Agglomerative Agent Memory System for Long-Term Conversations,https://arxiv.org/abs/2606.29778,True,core,execution_state,model,retrieval_access;representation_organization;update_forgetting;evaluation_diagnosis;systems_efficiency,,4,agent-memory
 2606.29788,2026-06-29,recent-2026-Apr-Jul,MemLeak: Diagnosing Information Leaks in Multimodal Agent Memory,https://arxiv.org/abs/2606.29788,True,core,update_forgetting,model,representation_organization;update_forgetting;evaluation_diagnosis;security_privacy_trust;multimodal_embodied_gui,benchmark_diagnosis;execution_multimodal;security_governance,7,agent-memory
 2606.29824,2026-06-29,recent-2026-Apr-Jul,Neural Procedural Memory: Empowering LLM Agents with Implicit Activation Steering,https://arxiv.org/abs/2606.29824,True,core,update_forgetting,model,retrieval_access;representation_organization;experience_skill_learning;evaluation_diagnosis;multimodal_embodied_gui,experience_skill_policy,5,agent-memory;procedural-memory
 2606.29914,2026-06-29,recent-2026-Apr-Jul,MemDelta: Controlled Baselines and Hidden Confounds in Agent Memory Evaluation,https://arxiv.org/abs/2606.29914,True,supporting,security_privacy,model,retrieval_access;update_forgetting;evaluation_diagnosis;systems_efficiency,benchmark_diagnosis,7,agent-memory;memory-evaluation
 2606.29961,2026-06-29,recent-2026-Apr-Jul,DuoMem: Towards Capable On-Device Memory Agents via Dual-Space Distillation,https://arxiv.org/abs/2606.29961,True,core,update_forgetting,model,experience_skill_learning;evaluation_diagnosis;multimodal_embodied_gui,experience_skill_policy;systems_efficiency,5,agent-memory
 2606.30296,2026-06-29,recent-2026-Apr-Jul,ManimAgent: Self-Evolving Multimodal Agents for Visual Education,https://arxiv.org/abs/2606.30296,False,core,execution_state,model,retrieval_access;update_forgetting;experience_skill_learning;evaluation_diagnosis;multimodal_embodied_gui;systems_efficiency,experience_skill_policy;execution_multimodal,6,episodic-memory
 2606.30566,2026-06-29,recent-2026-Apr-Jul,Forensic Trajectory Signatures for Agent Memory Poisoning Detection,https://arxiv.org/abs/2606.30566,True,core,update_forgetting,model,retrieval_access;execution_state;evaluation_diagnosis;security_privacy_trust;multimodal_embodied_gui,security_governance,4,agent-memory;long-term-memory
 2606.30639,2026-06-29,recent-2026-Apr-Jul,Self-Evolving World Models for LLM Agent Planning,https://arxiv.org/abs/2606.30639,False,core,update_forgetting,model,retrieval_access;execution_state;experience_skill_learning;evaluation_diagnosis,experience_skill_policy,4,episodic-memory
 2606.31046,2026-06-30,recent-2026-Apr-Jul,OpenLife: Toward Open-World Artificial Life with Autonomous LLM Agents,https://arxiv.org/abs/2606.31046,False,core,experience_learning,model,retrieval_access;representation_organization;experience_skill_learning;evaluation_diagnosis;systems_efficiency,,2,long-term-memory
 2606.31612,2026-06-30,recent-2026-Apr-Jul,What Memory Do GUI Agents Really Need? From Passive Records to Active Task-Driving States,https://arxiv.org/abs/2606.31612,True,core,update_forgetting,model,retrieval_access;update_forgetting;execution_state;experience_skill_learning;evaluation_diagnosis;multimodal_embodied_gui;systems_efficiency,benchmark_diagnosis;execution_multimodal,3,agent-memory
 2607.02579,2026-06-30,recent-2026-Apr-Jul,When Not to Write Memory: Governing False Promotion from Correlated Agent Traces,https://arxiv.org/abs/2607.02579,True,core,update_forgetting,model,retrieval_access;update_forgetting;evaluation_diagnosis;security_privacy_trust;systems_efficiency,update_forgetting;security_governance,4,agent-memory
 2607.00454,2026-07-01,recent-2026-Apr-Jul,Agri-SAGE: Simulation-Grounded Multi-Agent LLM for Context-Aware Agricultural Advisory Generation,https://arxiv.org/abs/2607.00454,False,core,update_forgetting,model,retrieval_access;evaluation_diagnosis;multi_agent_shared;multimodal_embodied_gui;systems_efficiency,personal_shared,6,episodic-memory
 2607.01047,2026-07-01,recent-2026-Apr-Jul,Conversable Complexity: Agentic LLM Collectives as Interpretable Substrates,https://arxiv.org/abs/2607.01047,False,core,execution_state,model,retrieval_access;experience_skill_learning;evaluation_diagnosis,personal_shared,3,long-term-memory
 2607.01071,2026-07-01,recent-2026-Apr-Jul,MemSyco-Bench: Benchmarking Sycophancy in Agent Memory,https://arxiv.org/abs/2607.01071,True,core,execution_state,model,retrieval_access;update_forgetting;evaluation_diagnosis;systems_efficiency,benchmark_diagnosis,4,agent-memory;memory-evaluation
 2607.01523,2026-07-01,recent-2026-Apr-Jul,Multi-Head Recurrent Memory Agents,https://arxiv.org/abs/2607.01523,True,core,update_forgetting,model,update_forgetting;evaluation_diagnosis;security_privacy_trust;systems_efficiency,,7,agent-memory
 2607.01709,2026-07-02,recent-2026-Apr-Jul,COMFYCLAW: Self-Evolving Skill Harnesses for Image Generation Workflows,https://arxiv.org/abs/2607.01709,False,core,update_forgetting,model,retrieval_access;experience_skill_learning;evaluation_diagnosis;multimodal_embodied_gui;systems_efficiency,experience_skill_policy,5,agent-memory
 2607.01916,2026-07-02,recent-2026-Apr-Jul,ContextSniper: AntTrail's Token-Efficient Code Memory for Repository-Level Program Repair,https://arxiv.org/abs/2607.01916,True,core,update_forgetting,model,retrieval_access;representation_organization;evaluation_diagnosis;systems_efficiency,systems_efficiency,6,agent-memory
 2607.01935,2026-07-02,recent-2026-Apr-Jul,A-TMA: Decoupling State-Aware Memory Failures in Long-Term Agent Memory,https://arxiv.org/abs/2607.01935,True,core,update_forgetting,model,retrieval_access;update_forgetting;evaluation_diagnosis,,3,agent-memory;long-term-memory
 2607.03726,2026-07-04,recent-2026-Apr-Jul,SelfMem: Self-Optimizing Memory for AI Agents,https://arxiv.org/abs/2607.03726,True,core,update_forgetting,model,retrieval_access;execution_state;experience_skill_learning;evaluation_diagnosis;multimodal_embodied_gui;systems_efficiency,,7,agent-memory
 2607.04089,2026-07-05,recent-2026-Apr-Jul,PLACEMEM: Toward a Compute-Aware Memory Plane for Lifelong Agents,https://arxiv.org/abs/2607.04089,True,core,update_forgetting,model,retrieval_access;update_forgetting;evaluation_diagnosis;security_privacy_trust;systems_efficiency,,2,agent-memory
 2607.04391,2026-07-05,recent-2026-Apr-Jul,Memory-Orchestrated Semantic System (MOSS): An Auditable Agentic Memory Architecture,https://arxiv.org/abs/2607.04391,True,core,update_forgetting,model,retrieval_access;representation_organization;systems_efficiency,,0,agent-memory
 2607.05029,2026-07-06,recent-2026-Apr-Jul,Your Agent's Memories Are Not Its Own: Forged Reasoning Attacks on LLM Agent Memory and Defenses,https://arxiv.org/abs/2607.05029,True,core,update_forgetting,model,retrieval_access;evaluation_diagnosis;security_privacy_trust,security_governance,6,agent-memory;long-term-memory
 2607.05577,2026-07-06,recent-2026-Apr-Jul,Narrative World Model: Narratology-Grounded Writer Memory for Long-Form Fiction,https://arxiv.org/abs/2607.05577,True,core,update_forgetting,model,retrieval_access;representation_organization;evaluation_diagnosis,,5,agent-memory
 2607.06595,2026-07-06,recent-2026-Apr-Jul,When Agents Remember Too Much: Memory Poisoning Attacks on Large Language Model Agents,https://arxiv.org/abs/2607.06595,True,core,update_forgetting,model,retrieval_access;execution_state;security_privacy_trust;systems_efficiency,security_governance,4,agent-memory;long-term-memory
 2607.06195,2026-07-07,recent-2026-Apr-Jul,LogicHunter: Testing LLM Agent Frameworks with an Agentic Oracle,https://arxiv.org/abs/2607.06195,False,core,update_forgetting,model,retrieval_access;representation_organization;evaluation_diagnosis;security_privacy_trust;multimodal_embodied_gui,,7,memory-evaluation
 2607.07108,2026-07-08,recent-2026-Apr-Jul,Seeing and Reflecting: Multimodal Memory-Enhanced Agent Collaboration for Recommendation,https://arxiv.org/abs/2607.07108,True,core,update_forgetting,model,representation_organization;update_forgetting;experience_skill_learning;multimodal_embodied_gui,execution_multimodal,5,agent-memory
 2607.08032,2026-07-09,recent-2026-Apr-Jul,"What to Keep, What to Forget: A Rate--Distortion View of Memory Compaction in LLMs and Agents",https://arxiv.org/abs/2607.08032,True,core,update_forgetting,model,retrieval_access;update_forgetting;evaluation_diagnosis;systems_efficiency,update_forgetting,5,agent-memory
 2607.08716,2026-07-09,recent-2026-Apr-Jul,Remember When It Matters: Proactive Memory Agent for Long-Horizon Agents,https://arxiv.org/abs/2607.08716,True,core,update_forgetting,model,retrieval_access;representation_organization;execution_state;evaluation_diagnosis;multimodal_embodied_gui,execution_multimodal,6,agent-memory
@@ -0,0 +1,66 @@
 # Agent Memory field audit
 status: completed
 date: 2026-07-10
 owner: Codex + user
 ## 触发原因
 12 篇锚点讲义被用户判定为无用。问题不是表达，而是样本太少、没有时间对照、没有拆实验归因，也没有解释大量论文在共同解决什么。
 ## 研究问题
 1. 最近六个月相对前六个月，哪些问题出现了可比较的能力进步？
 2. 哪些只是新架构、新数据集或更弱 baseline 下的表面增益？
 3. 静态召回、更新、执行状态、经验复用、写入治理和安全分别卡在哪里？
 4. 一个 memory 设计需要怎样的实验才能支持因果主张？
 ## 数据修正
 - 原论文采集器每个 query 默认只取最新 120 篇。2026-05 到 2026-07 的 `agent memory` 结果已经超过上限，因此旧库的月份分布有截断偏差，不能用于半年趋势。
 - 新工具 `tools/research/search_arxiv_memory.py` 使用分页和 5 组查询，覆盖 2025-01-01 至 2026-07-10，得到 516 个唯一候选。
 - 词法上界为 379 篇标题中心论文；最近六个月 292 篇，前六个月 61 篇。
 - 本地自动 `memory` 标签实际对应 293 篇论文，不是此前记录的 296；三条差异来自非论文条目。
 ## 流程
 1. 用分页 arXiv API 建立可复跑的候选池和摘要数据。
 2. 用本地 `ChatGPT-5.6:Terra` 做 516 篇 recall-oriented 初筛，只作为 triage，不承担结论。
 3. 用确定性词法规则生成时间窗口、标题议程和 evidence-cue 统计。
 4. 下载 32 篇 PDF 并转成文本，定位 introduction、主结果、消融、成本和 limitation。
 5. 只在同 benchmark、同论文受控设置或明确固定变量的研究中判断能力变化。
 6. 把正结果、负结果、成本和不能外推的范围同时写入证据账本。
 ## 模型分工
 - Ollama 小模型只做摘要级 relevance/problem triage。请求不指定 `num_ctx`，不发送 `keep_alive`。
 - 曾用 `ChatGPT-5.6:Sol` 测试细粒度抽取，约 3 分钟才处理 12 篇，未继续跑 293 篇；部分结果保存在 `memory-classified.json`，不参与领域结论。
 - 规则程序负责可复跑的日期、数量、关键词和窗口统计。
 - Codex 负责全文实验核验、跨论文冲突、归因和最终判断。
 - 本轮没有中控 GPT-5.5 的 API 信息，因此没有把它写成已执行步骤。
 ## 产物
 - `data/research/agent-memory-expanded.json`: 516 篇分页候选及摘要。
 - `data/research/agent-memory-screened.json`: 摘要级初筛结果。
 - `data/research/memory-corpus.json`: 本地 293 篇 `memory` 标签论文的完整摘要。
 - `data/research/memory-classified.json`: Sol 细粒度抽取的 12/293 篇中止样本。
 - `data/research/memory-landscape.json`: 时间窗口、方向和 evidence-cue 统计。
 - `data/research/memory-paper-ledger.csv`: 每篇候选的可过滤账本。
 - `research/memory/findings.md`: 直接回答问题的研究结论。
 - `research/memory/evidence-ledger.md`: 全文证据和边界。
 ## 有效经验
 - 论文数量不能作为进步指标；必须找相同任务或单变量控制。
 - 先找强负结果和 baseline audit，再读新架构，能快速发现假增益。
 - 摘要可以筛选“可能有证据”的论文，不能支持成本、公平性和失败边界。
 - “memory”必须拆成 write、store、retrieve、resolve、use、act 六个阶段，否则论文结果无法比较。
 - 同一个系统要分开报告静态 QA、动态更新和 agentic action；三者不能共享一个总分结论。
 ## 仍有限制
 - 516 是查询并集，自动初筛明显偏向高召回；379 是词法上界，不是人工确认核心数。
 - 32 篇全文仍不是 379 篇的全量系统综述，但已经覆盖关键 benchmark、受控反例和近期正结果。
 - 最近几个月的大量材料是未独立复现的 arXiv preprint。
 - 论文之间的模型和 harness 快速变化，跨论文只做问题趋势判断，不做绝对分数排行榜。
@@ -1,6 +1,6 @@
 # KC-001：Agent Memory 主题卷宗试验
-status: running
+status: failed
 date: 2026-07-10
 owner: Codex + user
@@ -23,7 +23,7 @@ owner: Codex + user
 - 本地 Agent 论文库：974 篇。
 - 时间范围：现有库覆盖的最近一年。
- 候选 memory 证据池：约 296 篇自动标签论文。
+- 候选 memory 证据池：293 篇自动标签论文。此前的 296 把非论文条目算了进去，现已纠正。
 - 本轮锚点：12 篇原始论文。
 ## 锚点选择规则
@@ -123,6 +123,20 @@ EvoMemBench、AutoMEM 诊断和预算研究对认知的贡献不亚于新方法
 ## 目前的失败与不足
 ### 用户判定：失败
 用户的反馈是“读的论文太少”“认知没有提高”“不是要听课，而是一起研究”。这个判定成立。本轮产物虽然结构完整，但没有回答最重要的研究问题：
 - memory 领域当前到底在解决哪些不同问题。
 - 半年前的失败今天有没有在可比实验中改善。
 - 200 多篇论文分别在做什么，而不是 12 篇锚点说了什么。
 - 哪些增益来自 memory，哪些来自模型、embedding、prompt、token 或额外调用。
 - 一个新设计需要什么证据才能说服别人。
 失败原因不是页面或讲义形式，而是研究采样和归因方式错误：12 篇锚点适合搭概念框架，不足以判断领域进展；跨论文总分不可比，却没有先建立受控比较和时间基线；输出形式像课程，把未充分验证的综合判断写成了结论。
 因此本试验不再以“让用户完成 45 分钟学习”为下一步。后续改为 field audit：完整分页采集、问题分层、同 benchmark 受控比较、全文证据账本和失败边界记录。新结果见 `research/memory/findings.md`。
 - 还没有真实用户学习数据，因此不能证明 45 分钟路径有效。
 - 12 篇锚点偏 2026 年最新工作，缺少对 2023-2025 经典系统的直接精读，历史演化主要由 survey 补足。
 - 定量结果来自不同模型、benchmark 和协议，只能逐篇解释，不能横向排序。
@@ -151,4 +165,3 @@ EvoMemBench、AutoMEM 诊断和预算研究对认知的贡献不亚于新方法
 - 两道设计题至少一道达到 2 分以上。
 - 一周后复习仍能恢复主要框架，而不是只记住论文名字。
 - 用户能指出至少一个讲义结论的证据边界。
@@ -0,0 +1,8 @@
 # Research
 这里保存跨论文、带时间对照和证据边界的研究结果。单篇资料仍归档在 `papers/`，研究结论不能只依赖摘要或论文标题。
 ## Current audits
 - [Memory findings](memory/findings.md): 最近六个月哪些问题真的有进展，哪些仍未解决。
 - [Memory evidence ledger](memory/evidence-ledger.md): 全文实验、反例、成本和不能外推的范围。
@@ -0,0 +1,37 @@
 # 证据账本
 `full` 表示核过实验表、关键消融或 limitation；`context` 表示全文已转文本并做章节定位，但不承担本轮核心结论。
 | Date | Paper | Depth | 可支持的判断 | 不能外推的部分 |
 | --- | --- | --- | --- | --- |
 | 2025-01 | [Zep, 2501.13956](https://arxiv.org/abs/2501.13956) | full | 长对话静态 QA 可用时间图谱大幅缩短上下文和延迟，并提升部分题型 | DMR 已饱和；single-session-assistant 反而下降；不是行动任务 |
 | 2025-02 | [A-MEM, 2502.12110](https://arxiv.org/abs/2502.12110) | full | 动态 linked-note 组织在 LoCoMo 上能优于若干早期 memory baseline | 方法、模型和 token 配置并非严格单变量控制 |
 | 2025-05 | [Experience-Following, 2505.16067](https://arxiv.org/abs/2505.16067) | full | 经验会影响后续相似任务，错误和错配经验会传播 | 不能证明某种通用写入或删除策略最优 |
 | 2025-07 | [MemoryAgentBench, 2507.05257](https://arxiv.org/abs/2507.05257) | full | retrieval、TTL、LRU、forgetting 是不同能力；复杂更新和全局综合仍弱 | 数据集转换和统一 prompt 不能覆盖真实 agent loop |
 | 2026-02 | [MemoryArena, 2602.16313](https://arxiv.org/abs/2602.16313) | full | 通用 memory/RAG 在相互依赖的多 session 行动任务上成功率低，额外 memory 延迟更高 | benchmark 本身很难，失败同时含 policy/reasoning 问题 |
 | 2026-03 | [Retrieval vs Utilization, 2603.02473](https://arxiv.org/abs/2603.02473) | full | LoCoMo 上 retrieval 方法差异大于 write strategy，raw RAG 是强 baseline | 单模型、单 benchmark、固定 top-k=5 |
 | 2026-05 | [Cross-Component Interference, 2605.05716](https://arxiv.org/abs/2605.05716) | full | memory 与其他 scaffold 的边际作用可为负，必须做 subset ablation | 只有 HotpotQA/GSM8K；tool 组件含协议因素 |
 | 2026-05 | [EvoMemBench, 2605.18421](https://arxiv.org/abs/2605.18421) | full | 15 种方法间不存在跨能力通用最优 memory form | 统一 benchmark 仍不能消除所有实现质量差异 |
 | 2026-05 | [DeMem, 2605.10870](https://arxiv.org/abs/2605.10870) | full | 决策充分状态比描述相似性更适合行动记忆；LoCoMo 描述相似性相关很弱 | 主要是自定义决策任务和理论建模 |
 | 2026-05 | [MemQ, 2605.08374](https://arxiv.org/abs/2605.08374) | full | provenance DAG credit assignment 在多步任务有增益，单步接近无增益 | 没有处理删除和长期错误积累 |
 | 2026-06 | [AutoMEM, 2606.04315](https://arxiv.org/abs/2606.04315) | full | 跨场景无通用赢家；索引会产生 schema loss，passive retrieval 会产生 retrieval loss | 新增 DCI harness 与已有系统的工程成熟度不同 |
 | 2026-06 | [MAGE, 2606.06090](https://arxiv.org/abs/2606.06090) | full | MemoryArena 同设置下，active execution-state tree 同时提高 SR 并减少 token | 单团队结果；绝对成功率在 Travel 仍只有 0.1519 |
 | 2026-06 | [Infini Memory, 2606.10677](https://arxiv.org/abs/2606.10677) | full | topic document 的 maintenance 对结果贡献明显，memory 维护不应被当后台清理 | 系统组件较多，维护和检索仍依赖 LLM |
 | 2026-06 | [HORMA, 2606.11680](https://arxiv.org/abs/2606.11680) | full | organize 与 retrieve 可拆分；高能力 manager 比大 retriever 更关键 | 结论依赖其层级 memory 和任务设置 |
 | 2026-06 | [Online Budget Study, 2606.15017](https://arxiv.org/abs/2606.15017) | full | budget-matched 时，vanilla agent 在 WebArena 多模型汇总上可胜 online skill/memory | 只研究 online augmentation，不覆盖离线学习 |
 | 2026-06 | [TrustMem, 2606.25161](https://arxiv.org/abs/2606.25161) | full | transition-level verifier 和排名训练可降低 omission/corruption/hallucination | verifier 仍是模型；训练、judge 和跨域泛化需要独立核验 |
 | 2026-06 | [Supersede, 2606.27472](https://arxiv.org/abs/2606.27472) | full | 当前值维护是独立瓶颈，更强模型和更大 memory 不足；训练信号能推动它 | 训练只有 Qwen2.5-3B 单次运行，16.7% 远非解决 |
 | 2026-06 | [MemDelta, 2606.29914](https://arxiv.org/abs/2606.29914) | full | embedding、model behavior 和 write cost 足以翻转架构结论 | Mem0 只在 88 题、2/6 题型上做 matched comparison |
 | 2026-06 | [ATMem, 2606.31612](https://arxiv.org/abs/2606.31612) | full | GUI agent 的主动任务状态和选择性调用能提高结果 | 128xH20 训练；更难 DataScope 仍接近失败 |
 | 2026-06 | [GovMem, 2607.02579](https://arxiv.org/abs/2607.02579) | full | 自动写入必须检查独立证据、反例、scope 和 provenance；真实轨迹暴露 false promotion | 外部 133 条是 stress slice，安全自动覆盖率为 0 |
 | 2026-07 | [A-TMA, 2607.01935](https://arxiv.org/abs/2607.01935) | full | current/historical/transition 角色能缓解 ghost memory | 收益 host-dependent，LoCoMo 部分指标不一致 |
 | 2026-07 | [AgenticSTS, 2607.02255](https://arxiv.org/abs/2607.02255) | full | typed memory contract 让技能/episode 可消融和复现 | 3/10 到 6/10 不显著；缺 same-code accumulating-context baseline |
 | 2026-07 | [FARMA, 2607.05029](https://arxiv.org/abs/2607.05029) | full | 伪造推理记录是不同于事实投毒的写入攻击面 | 50-trial 自定义攻击；0% ASR 防御不等于开放环境稳健 |
 ## 全文池
 本轮全文池共 32 篇：
 `2501.13956`, `2502.12110`, `2505.16067`, `2507.02259`, `2507.05257`, `2507.07957`, `2511.20857`, `2512.18746`, `2601.01885`, `2602.16313`, `2602.19320`, `2603.02473`, `2603.07670`, `2603.15666`, `2605.05716`, `2605.08374`, `2605.10870`, `2605.14421`, `2605.18421`, `2606.04315`, `2606.06090`, `2606.10677`, `2606.11680`, `2606.15017`, `2606.25161`, `2606.27472`, `2606.29914`, `2606.31612`, `2607.01935`, `2607.02255`, `2607.02579`, `2607.05029`。
 其中没有进入上表核心论证的 9 篇暂标为 `context`，不把“下载过 PDF”写成“已得到结论”。
@@ -0,0 +1,105 @@
 # 先说结论
 截至 2026-07-10，这轮可复核的材料是：5 组分页 arXiv 查询得到 516 篇候选；其中 379 篇是标题含 memory/remember/forget 等词、正文处于 LLM Agent 语境的词法上界；最近六个月有 292 篇，前六个月只有 61 篇。论文数量增加了约 4.8 倍，但这不能解释成能力增加了 4.8 倍。
 本轮下载并转成可检索文本的全文是 32 篇。下面的判断主要依靠其中 23 篇的实验表、消融和限制章节，不用标题和摘要投票。
 能站住的判断只有四条：
 1. **真正进步最大的是诊断和评测，不是通用记忆能力。** 研究开始把写入、检索、使用、当前状态、行动结果和成本拆开测，也开始暴露 embedding、基础模型、上下文长度和额外 LLM 调用造成的假增益。
 2. **静态对话事实召回已经是可用工程问题，但没有证据证明复杂 memory 普遍优于强 RAG 或完整上下文。** 在一个 3x3 控制实验中，换检索方法造成 14-23 个准确率点差异，换写入方法只有 3-8 个点；原始文本 RAG 还击败了两种有损写入。
 3. **最近半年唯一比较像“能力进步”的，是把长任务记忆从资料库改成执行状态。** MAGE 在同一篇论文、同一 backbone、同一 MemoryArena 协议下，平均比长上下文高 7.8 个成功率点，同时少用 32.9%-71.4% token。它证明了任务状态表示是有效方向，但还只是一个基准上的单篇结果。
 4. **更新、遗忘、经验复用和自动写入仍未解决。** 更强模型和更大 memory 都不能稳定消除旧事实污染；经验可能传播错误；保守写入可以降低错误晋升，却把代价转成大量人工复核。
 ## 具体问题有没有变好
 | 问题 | 六个月前能做到什么 | 最近证据 | 判断 |
 | --- | --- | --- | --- |
 | 静态事实召回 | 图、note、摘要等结构在 LoCoMo/LongMemEval 上可以优于较弱 baseline；Zep 用约 1.6k 上下文替代 115k，LongMemEval 从 55.4% 到 63.8% 或 60.2% 到 71.2% | 控制实验表明检索器比复杂写入更重要；MemDelta 中 verbatim RAG 47.2% 与 full context 49.8% 无显著差异 | 工程上部分成熟；“新结构更聪明”没有普遍证据 |
 | 长范围综合和 test-time learning | top-k retrieval 经常拿不到全局信息；长上下文在预算够时领先 | compute-matched 结果显示 book summarization 只有拿到接近完整信息后才明显上升，架构差异次要 | 没解决，瓶颈是信息覆盖阈值 |
 | 当前值、更新和选择性遗忘 | MemoryAgentBench 的复杂多跳更新普遍很低，简单 overwrite prompt 无法修复 | Supersede 中 bounded memory 相对 full context 分别为 63% vs 82%、64% vs 91%、77% vs 92%；历史增长 24 倍后 68% 跌到 28%，memory 同比扩大仍是 28% | 没解决；现在只是把问题隔离得更清楚 |
 | 长任务执行状态 | 通用 memory/RAG 在 MemoryArena 上平均成功率约 0.12-0.15/0.17-0.23，Group Travel 几乎全为 0 | MAGE 把 memory 变成可 Grow/Compress/Maintain/Revise 的状态树，平均 +7.8pp SR；GUI 的 ATMem 也显示主动任务状态有效 | 有真实但任务特定的进展，需要独立复现 |
 | 经验和技能复用 | 相似任务会复用相似经验，但错误经验也会让输出沿相同方向出错 | 32k+ full-factorial 评测中 memory 对 HotpotQA 为负、对 GSM8K 略正；AgenticSTS 的技能层 3/10 到 6/10，Fisher p 约 0.37 | 没有通用增益，强依赖任务、模型和 scaffold |
 | 自动写入和压缩 | 主要优化“写得短、写得结构化” | TrustMem 开始监督每次 transition；GovMem 把写入改成证据治理，但内部 direct recall 降到 0.448、review burden 升到 0.692，外部 133 条高影响候选中安全自动晋升为 0 | 研究方向变对了，自动写入仍不可靠 |
 | 安全和来源 | 主要研究检索语料投毒 | 新工作覆盖持久污染、伪造推理、共享 memory 传播、provenance 和 false promotion | 威胁模型进步明显；防御多为自定义攻击上的单篇结果 |
 | 评测可信度 | 常见做法是端到端比一个总分，模型、embedding、token 和写入成本都不同 | MemDelta 只换 embedding 就 +6.2pp，并把 Mem0 相对 RAG 的 +11pp 变成 -1.2pp；相近效果下写入成本约 50 倍 | 这是最近半年最确定的进步 |
 ## 这些论文实际上在做什么
 最近六个月 292 篇“标题中心”候选中，按标题声明的贡献做多标签统计：
 | 方向 | 篇数 | 实际工作 |
 | --- | ---: | --- |
 | retrieval / representation | 65 | 图、层级、索引、rerank、topic document、typed memory |
 | experience / skill / learned policy | 65 | 经验提炼、技能库、RL 写入/检索/遗忘策略、自演化 |
 | execution / multimodal | 58 | GUI、embodied、video、spatial、long-horizon task state |
 | benchmark / diagnosis | 56 | 新 benchmark、stage-wise probe、跨场景和受控 baseline |
 | systems / efficiency | 37 | token、延迟、预算、压缩、端侧和 serving |
 | security / governance | 32 | poisoning、privacy、provenance、写入 gate、审计 |
 | personalization / shared memory | 27 | 用户长期状态、多 Agent 共享和协作 memory |
 | update / forgetting 明示在标题 | 22 | 冲突、时序、过期、遗忘、写入和生命周期 |
 这些数字可重叠，而且只是研究议程，不是成果计数。数量暴涨主要表示大家把“memory”扩展成了八九个不同问题；不是 292 个方法都在共同推进同一个分数。
 ## 哪些结果最值得信
 ### 1. 静态 QA 先做强 retrieval，不要先发明 memory 架构
 [Diagnosing Retrieval vs. Utilization](https://arxiv.org/abs/2603.02473) 在 LoCoMo 1540 题上固定 answer model 和 prompt，交叉三种写法与三种检索法。Basic RAG 在 cosine/hybrid 下为 77.9%/81.1%，fact extraction 为 72.2%/77.3%，episode summary 为 70.1%/73.3%；retrieval precision 与最终准确率相关系数为 0.98。这个结论只适用于静态对话 QA 和固定 top-k=5，但它足以否定“复杂写入天然更好”。
 [MemDelta](https://arxiv.org/abs/2606.29914) 更进一步：500 题上 cloud embedding RAG 为 53.4%，MiniLM RAG 为 47.2%，完整上下文为 49.8%。在可做 matched comparison 的 88 题上，Mem0 72.7%，cloud RAG 73.9%，p=1.0 且置信区间很宽；Mem0 需要 1000+ LLM 调用、约 120 分钟和 $0.50+ 写入成本，RAG 是 0 次 LLM 调用、约 60 秒和 $0.01。它没有证明 memory 无用，只证明当前很多“架构增益”无法归因。
 ### 2. 更新问题不是检索一个最新 chunk 那么简单
 [MemoryAgentBench](https://arxiv.org/abs/2507.05257) 把能力拆成准确检索、test-time learning、长范围理解和选择性遗忘。GPT-4.1-mini 的默认 single-hop/multi-hop 更新为 36%/5%；强制总选最新后为 40%/4%，保守 overwrite 为 28%/4%。prompt 没有传播多跳依赖上的更新。
 [Supersede](https://arxiv.org/abs/2606.27472) 把 full context 与 bounded self-maintained memory 配对比较，证明维护过程本身丢失当前值。其 GRPO 训练把 Qwen2.5-3B 从 9.0% 提到 16.7%，但只有一次训练、绝对结果仍很低。合理表述是“这个失败可被训练信号推动”，不是“已经补上差距”。
 [A-TMA](https://arxiv.org/abs/2607.01935) 用 current/historical/transition 角色缓解 ghost memory：Graphiti/Zep 的冲突准确率从 0.480 到 0.720，LoCoMo temporal F1 从 0.0295 到 0.1705；作者也报告 host-dependent 和非一致增益。它支持显式状态角色，不支持通用解决方案。
 ### 3. 行动任务需要的是最小充分状态，不是相似历史
 [MemoryArena](https://arxiv.org/abs/2602.16313) 把 memory 放回 agent-environment loop。通用 memory 系统在 Bundled Shopping 全部 0 成功，在 Group Travel 全部 0；RAG 的总体平均成功率最高也只有 0.23。作者观察到 representation mismatch、training mismatch 和额外延迟。
 [MAGE](https://arxiv.org/abs/2606.06090) 是目前最值得继续追的正结果。它在同一 Qwen3.6-27B 设置中，Shopping 0.3333 到 0.3933，Travel 0.0519 到 0.1519，Web Search 0.4842 到 0.5656；移除 Compress/Maintain/Revise 都会下降。还缺跨团队复现、更多模型和更难任务上的绝对成功率。
 [ATMem](https://arxiv.org/abs/2606.31612) 在 AndroidWorld 8B 上从基础 47.6，经 SFT/GRPO/结构奖励到 76.6，但训练用了 128 张 H20；在更难的 DataScope 分层任务上绝对成功率仍只有 6.2/6.2/3.1。它证明 active state 有价值，也同时说明代价和剩余差距巨大。
 ### 4. 写入的本质正在从总结变成决策和治理
 [GovMem](https://arxiv.org/abs/2607.02579) 的内部 synthetic 结果把 false promotion 从 0.597 降到 0.040、recall 保持 0.960；但真实项目子集依靠 0.692 review burden，外部 133 条高影响 coding-agent 候选最终没有一条可安全自动晋升。这个负结果比一个漂亮总分更有用：重复出现、测试日志和 Agent 自己的叙述都不能单独构成可复用知识证据。
 [Cross-Component Interference](https://arxiv.org/abs/2605.05716) 在 32 个组件子集、32k+ 次评测中发现，HotpotQA 上 tool-only F1 0.233，高于 all-in 的 0.177；最优 proper subset 在所有测试设置中匹配或超过 all-in。它要求每个 memory 设计都必须做 no-memory 和组件组合消融。
 ## 怎么证明一个 memory 设计不是画图
 一份能说服人的实验至少要同时满足：
 1. 先说工作负载：信息是否超窗、是否会更新、是否跨任务、是否影响行动、错误写入会不会传播。
 2. 指定瓶颈阶段：write preservation、retrieval recall、state resolution、utilization、action outcome 只能选清楚，不能只报端到端总分。
 3. 必须有 full context、verbatim RAG、no memory 三个 baseline；做技能复用时还要有 raw trajectory replay。
 4. actor model、embedding、chunking、top-k、prompt、可见信息和 token/调用预算固定，只改一个组件。
 5. 报预处理时间、LLM 调用、推理 token、延迟和存储；不能只报“检索后上下文更短”。
 6. 做 harmful-memory、stale-memory 和 no-memory 消融；memory 可能是负贡献。
 7. 至少跨两个任务类型、两个模型、多个 seed；否则结论只能写成“在该 harness 中有效”。
 可以把论证压成一句话：**在工作负载 W 中，阶段 B 的失败占主要误差；固定其余变量后，机制 C 把结果从 X 提到 Y，额外成本是 Z，且在边界 Q 外不成立。** 架构图只能解释 C，不能替代 W、B、X、Y、Z 和 Q。
 ## 对我们自己的直接影响
 现在不应该先做一个“统一 Agent Memory 知识图谱”。更合理的最小系统是四个彼此独立、都保留原始证据的平面：
 - `archive`：原始论文、对话、轨迹和版本，只追加，不让摘要覆盖证据。
 - `current state`：明确哪些事实和任务状态当前有效，保留 supersession 链。
 - `skills`：只保存有任务结果和适用范围支撑的程序性经验，允许回滚。
 - `governance`：记录来源、反例、写入决定、评测结果和人工复核。
 检索先从 raw archive + 强 baseline 做起。只有在确定是 retrieval、state resolution、execution state 或 skill transfer 的哪一种失败后，才增加对应结构。否则我们很容易复现这半年论文里最常见的问题：系统越来越复杂，但不知道增益来自哪里。
 ## 材料边界
 - 516 是 5 组 arXiv 查询的并集，不是人工确认的 516 篇核心论文。
 - 379 和 292 是保守词法筛选后的上界，仍可能有误收；标题主题统计可重叠。
 - 32 篇做了全文下载和章节定位，23 篇进入证据账本；其余论文目前只用于背景和冲突检索。
 - 大多数 2026 工作仍是 arXiv preprint，独立复现很少。这里把“同论文内受控增益”和“跨论文已建立共识”严格分开。
 - 跨 benchmark 的绝对分数不能直接排名；这里只比较同协议内结果，或明确写出不可比原因。
@@ -0,0 +1,285 @@
 #!/usr/bin/env python3
 """Build auditable landscape counts and evidence candidates for Agent Memory."""
 from __future__ import annotations
 import argparse
 import csv
 import json
 import re
 from collections import Counter, defaultdict
 from datetime import date
 from pathlib import Path
 ROOT = Path(__file__).resolve().parents[2]
 THEMES = {
    "retrieval_access": r"retriev|recall|search|query|evidence access|memory navigation",
    "representation_organization": r"knowledge graph|graph memory|hierarch|structure|representation|index|database|topic document|ontology",
    "update_forgetting": r"update|forget|consolidat|supersed|stale|conflict|write path|write-path|memory admission|retention",
    "execution_state": r"execution state|task state|workflow state|belief state|long-horizon|trajectory|subgoal|planning",
    "experience_skill_learning": r"experience|procedural|skill|reflect|self-evolv|test-time learn|continual learn|reinforcement learn|\brl\b",
    "evaluation_diagnosis": r"benchmark|evaluat|diagnos|metric|ablation|empirical study|controlled",
    "security_privacy_trust": r"attack|poison|security|privacy|trust|provenance|govern|leak|integrity|adversarial",
    "multi_agent_shared": r"multi-agent|multiagent|shared memory|collective memory|collaborative memory",
    "multimodal_embodied_gui": r"multimodal|visual|video|robot|embodied|gui|mobile|web agent|navigation",
    "systems_efficiency": r"token|latency|cost|efficient|compression|serving|storage|throughput|budget",
 }
 # Title-only labels describe the paper's declared contribution more conservatively
 # than abstract keywords such as "evaluate", which occur in almost every paper.
 TITLE_THEMES = {
    "benchmark_diagnosis": r"benchmark|evaluat|diagnos|empirical|survey|taxonomy|anatomy|rethinking|what .* need|how far|study of|analysis",
    "retrieval_representation": r"retriev|recall|search|graph|hierarch|structure|index|database|ontology|organization|representation|rag\b|topic document",
    "update_forgetting": r"update|forget|consolidat|stale|supersed|decay|write memory|not to write|promotion|temporal conflict|belief revision|lifecycle",
    "experience_skill_policy": r"experience|skill|procedural|learn|evolv|adapt|reinforcement|\brl\b|credit assignment|distill",
    "execution_multimodal": r"execution state|task-driving state|task state|long-horizon|gui|embodied|robot|navigation|web agent|multimodal|video|spatial",
    "security_governance": r"attack|poison|security|privacy|trust|govern|provenance|integrity|leak|hijack|vulnerab|unlearn|false promotion|contamination",
    "systems_efficiency": r"efficient|cost|token|latency|serving|on-device|edge|compress|scalable|storage|runtime|budget|wire format",
    "personal_shared": r"personali|user-aware|user-centric|multi-agent|multiagent|shared memory|collaborative|collective",
 }
 WINDOWS = {
    "previous_six_months": ("2025-07-09", "2026-01-10"),
    "recent_six_months": ("2026-01-10", "2026-07-11"),
 }
 CENTRAL_TITLE = re.compile(r"memory|memori|remember|forget|stateful", re.I)
 AGENT_CONTEXT = re.compile(r"agent|large language model|\bllm\b|foundation model", re.I)
 HARDWARE_ONLY = re.compile(
    r"gpu memory|memory bandwidth|memory footprint|memory-efficient|memory efficient|bounded memory agents on",
    re.I,
 )
 NUMERIC_RESULT = re.compile(r"(?:\d+(?:\.\d+)?\s*%|\+?\d+(?:\.\d+)?\s*(?:pp|points?|×|x))", re.I)
 def period(published: str) -> str:
    if published < "2025-07-01":
        return "2025-H1"
    if published < "2026-02-01":
        return "baseline-2025H2-Jan2026"
    if published < "2026-04-01":
        return "transition-2026-Feb-Mar"
    return "recent-2026-Apr-Jul"
 def central_by_text(record: dict) -> bool:
    title = record["title"]
    abstract = record["abstract"]
    return bool(
        CENTRAL_TITLE.search(title)
        and AGENT_CONTEXT.search(f"{title} {abstract}")
        and not HARDWARE_ONLY.search(title)
    )
 def theme_labels(record: dict) -> list[str]:
    text = f"{record['title']} {record['abstract']}".lower()
    return [name for name, pattern in THEMES.items() if re.search(pattern, text, re.I)]
 def title_theme_labels(record: dict) -> list[str]:
    title = record["title"].lower()
    return [name for name, pattern in TITLE_THEMES.items() if re.search(pattern, title, re.I)]
 def window_name(published: str) -> str | None:
    for name, (start, end) in WINDOWS.items():
        if start <= published < end:
            return name
    return None
 def evidence_score(record: dict) -> int:
    text = record["abstract"].lower()
    score = 0
    score += 2 if re.search(r"benchmark|evaluate|evaluation|experiment", text) else 0
    score += 2 if re.search(r"baseline|compare|outperform|versus|\bvs\.?\b", text) else 0
    score += 2 if NUMERIC_RESULT.search(text) else 0
    score += 1 if re.search(r"ablation|controlled|factorial|paired|significan", text) else 0
    score += 1 if re.search(r"code is|github|we release|open-source", text) else 0
    score += 1 if re.search(r"limitation|does not|not always|far from|fails?", text) else 0
    return score
 def write_json(path: Path, payload: dict) -> None:
    path.parent.mkdir(parents=True, exist_ok=True)
    path.write_text(json.dumps(payload, ensure_ascii=False, indent=2) + "\n", encoding="utf-8")
 def main() -> int:
    parser = argparse.ArgumentParser()
    parser.add_argument(
        "--input",
        type=Path,
        default=ROOT / "data" / "research" / "agent-memory-expanded.json",
    )
    parser.add_argument(
        "--screen",
        type=Path,
        default=ROOT / "data" / "research" / "agent-memory-screened.json",
    )
    parser.add_argument(
        "--output",
        type=Path,
        default=ROOT / "data" / "research" / "memory-landscape.json",
    )
    parser.add_argument(
        "--csv",
        type=Path,
        default=ROOT / "data" / "research" / "memory-paper-ledger.csv",
    )
    args = parser.parse_args()
    source = json.loads(args.input.read_text(encoding="utf-8"))
    screen = {}
    if args.screen.exists():
        screened = json.loads(args.screen.read_text(encoding="utf-8"))
        screen = {item["arxiv_id"]: item.get("screen") or {} for item in screened.get("papers", [])}
    records = []
    for paper in source["papers"]:
        screened = screen.get(paper["arxiv_id"])
        records.append(
            {
                **paper,
                "period": period(paper["published"]),
                "central_by_title": central_by_text(paper),
                "themes": theme_labels(paper),
                "title_themes": title_theme_labels(paper),
                "evidence_score": evidence_score(paper),
                "screen": screened,
            }
        )
    period_counts: defaultdict[str, Counter] = defaultdict(Counter)
    theme_counts: defaultdict[str, Counter] = defaultdict(Counter)
    window_counts: defaultdict[str, Counter] = defaultdict(Counter)
    title_theme_counts: defaultdict[str, Counter] = defaultdict(Counter)
    screen_counts: defaultdict[str, Counter] = defaultdict(Counter)
    for record in records:
        bucket = record["period"]
        period_counts[bucket]["candidates"] += 1
        if record["central_by_title"]:
            period_counts[bucket]["title_central"] += 1
        relevance = (record.get("screen") or {}).get("relevance")
        if relevance:
            period_counts[bucket][f"screen_{relevance}"] += 1
        for theme in record["themes"]:
            if record["central_by_title"]:
                theme_counts[bucket][theme] += 1
        window = window_name(record["published"])
        if window:
            window_counts[window]["candidates"] += 1
            if record["central_by_title"]:
                window_counts[window]["title_central"] += 1
                for theme in record["title_themes"]:
                    title_theme_counts[window][theme] += 1
        screened = record.get("screen") or {}
        if screened:
            screen_counts["relevance"][screened.get("relevance", "unknown")] += 1
            screen_counts["problem"][screened.get("problem", "unknown")] += 1
            screen_counts["source"][screened.get("source", "unknown")] += 1
    top_evidence = {}
    for theme in THEMES:
        candidates = [
            record
            for record in records
            if record["central_by_title"] and theme in record["themes"]
        ]
        candidates.sort(key=lambda item: (item["evidence_score"], item["published"]), reverse=True)
        top_evidence[theme] = [
            {
                "arxiv_id": item["arxiv_id"],
                "published": item["published"],
                "title": item["title"],
                "score": item["evidence_score"],
                "url": item["url"],
            }
            for item in candidates[:20]
        ]
    output = {
        "generated_at": date.today().isoformat(),
        "source_candidates": len(records),
        "title_central": sum(record["central_by_title"] for record in records),
        "screened": len(screen),
        "windows": {
            name: {
                "from_inclusive": start,
                "to_exclusive": end,
                **dict(window_counts[name]),
                "title_theme_counts": dict(title_theme_counts[name]),
            }
            for name, (start, end) in WINDOWS.items()
        },
        "period_counts": {key: dict(value) for key, value in period_counts.items()},
        "theme_counts_title_central": {key: dict(value) for key, value in theme_counts.items()},
        "screen_counts": {key: dict(value) for key, value in screen_counts.items()},
        "evidence_score_distribution": dict(
            Counter(record["evidence_score"] for record in records if record["central_by_title"])
        ),
        "top_evidence_candidates": top_evidence,
    }
    write_json(args.output, output)
    args.csv.parent.mkdir(parents=True, exist_ok=True)
    with args.csv.open("w", encoding="utf-8", newline="") as handle:
        writer = csv.DictWriter(
            handle,
            fieldnames=[
                "arxiv_id",
                "published",
                "period",
                "title",
                "url",
                "title_central",
                "screen_relevance",
                "screen_problem",
                "screen_source",
                "themes",
                "title_themes",
                "evidence_score",
                "matched_queries",
            ],
        )
        writer.writeheader()
        for record in records:
            screened = record.get("screen") or {}
            writer.writerow(
                {
                    "arxiv_id": record["arxiv_id"],
                    "published": record["published"],
                    "period": record["period"],
                    "title": record["title"],
                    "url": record["url"],
                    "title_central": record["central_by_title"],
                    "screen_relevance": screened.get("relevance", ""),
                    "screen_problem": screened.get("problem", ""),
                    "screen_source": screened.get("source", ""),
                    "themes": ";".join(record["themes"]),
                    "title_themes": ";".join(record["title_themes"]),
                    "evidence_score": record["evidence_score"],
                    "matched_queries": ";".join(record["matched_queries"]),
                }
            )
    print(
        json.dumps(
            {
                "source_candidates": output["source_candidates"],
                "title_central": output["title_central"],
                "screened": output["screened"],
                "windows": output["windows"],
                "screen_counts": output["screen_counts"],
            },
            ensure_ascii=False,
            indent=2,
        )
    )
    return 0
 if __name__ == "__main__":
    raise SystemExit(main())
@@ -0,0 +1,173 @@
 #!/usr/bin/env python3
 """Classify an Agent-memory abstract corpus into a research evidence ledger."""
 from __future__ import annotations
 import argparse
 import json
 import time
 import urllib.error
 import urllib.request
 from datetime import date
 from pathlib import Path
 ROOT = Path(__file__).resolve().parents[2]
 DEFAULT_OLLAMA_URL = "http://192.168.1.10:11434"
 SYSTEM_INSTRUCTIONS = """You are building an evidence ledger for a research review of Agent Memory.
 Classify each paper from title and abstract only. Be conservative. Do not infer unreported baselines,
 results, limitations, or artifacts. A paper is core only when memory itself is the main research object;
 supporting means memory is a substantial mechanism or measured failure; peripheral means memory is
 incidental or only an implementation detail.
 Return one JSON object with a `papers` array in the same order. Each item must contain:
 - arxiv_id: exact input id
 - relevance: core | supporting | peripheral
 - relevance_reason: <= 18 words
 - primary_problem: one of retention_context, retrieval_access, organization_representation,
  update_consolidation_forgetting, execution_state, experience_skill_learning,
  evaluation_measurement, security_trust_privacy, multi_agent_shared_memory,
  systems_efficiency, domain_application, theory_foundations, other
 - secondary_problems: up to 3 values from the same enum
 - research_role: method | benchmark | diagnostic | theory | survey | security | application | infrastructure
 - memory_object: facts | events | trajectories | skills | execution_state | model_internal |
  shared_state | multimodal | mixed | unclear
 - temporal_scope: in_episode | cross_episode | both | unclear
 - claimed_gap: one short sentence, or empty
 - mechanism: one short sentence, or empty
 - benchmarks: explicit names only, as an array
 - baselines: explicit names or baseline families only, as an array
 - reported_results: explicit numeric claims only, as an array of short strings
 - evidence_design: none | case_study | benchmark_comparison | controlled_ablation |
  formal_analysis | survey_synthesis
 - evidence_strength: none | weak | moderate | strong
 - failure_or_boundary: only a boundary explicitly visible in the abstract, or empty
 Use compact English. Output JSON only."""
 def write_json(path: Path, payload: dict) -> None:
    path.parent.mkdir(parents=True, exist_ok=True)
    tmp = path.with_suffix(path.suffix + ".tmp")
    tmp.write_text(json.dumps(payload, ensure_ascii=False, indent=2) + "\n", encoding="utf-8")
    tmp.replace(path)
 def ollama_generate(url: str, model: str, prompt: str, timeout: int, retries: int) -> dict:
    payload = {
        "model": model,
        "prompt": prompt,
        "stream": False,
        "format": "json",
        "options": {"temperature": 0, "num_predict": 6000},
    }
    request = urllib.request.Request(
        f"{url.rstrip('/')}/api/generate",
        data=json.dumps(payload).encode("utf-8"),
        headers={"Content-Type": "application/json"},
        method="POST",
    )
    for attempt in range(retries + 1):
        try:
            with urllib.request.urlopen(request, timeout=timeout) as response:
                body = json.loads(response.read().decode("utf-8"))
            return json.loads(body["response"])
        except (urllib.error.HTTPError, urllib.error.URLError, TimeoutError, json.JSONDecodeError, KeyError):
            if attempt >= retries:
                raise
            time.sleep(2 ** attempt)
    raise RuntimeError("unreachable retry state")
 def classify_batch(
    records: list[dict], url: str, model: str, timeout: int, retries: int
 ) -> list[dict]:
    compact = [
        {"arxiv_id": item["arxiv_id"], "title": item["title"], "abstract": item["abstract"]}
        for item in records
    ]
    prompt = f"{SYSTEM_INSTRUCTIONS}\n\nINPUT:\n{json.dumps(compact, ensure_ascii=False)}"
    result = ollama_generate(url, model, prompt, timeout=timeout, retries=retries)
    classified = result.get("papers") if isinstance(result, dict) else None
    if not isinstance(classified, list):
        raise ValueError("model response lacks papers array")
    expected = [item["arxiv_id"] for item in records]
    received = [str(item.get("arxiv_id")) for item in classified]
    if received != expected:
        raise ValueError(f"model returned unexpected ids: expected={expected}, received={received}")
    return classified
 def main() -> int:
    parser = argparse.ArgumentParser()
    parser.add_argument(
        "--input", type=Path, default=ROOT / "data" / "research" / "memory-corpus.json"
    )
    parser.add_argument(
        "--output", type=Path, default=ROOT / "data" / "research" / "memory-classified.json"
    )
    parser.add_argument("--ollama-url", default=DEFAULT_OLLAMA_URL)
    parser.add_argument("--model", default="ChatGPT-5.6:Sol")
    parser.add_argument("--batch-size", type=int, default=12)
    parser.add_argument("--timeout", type=int, default=300)
    parser.add_argument("--retries", type=int, default=2)
    parser.add_argument("--limit", type=int)
    parser.add_argument("--refresh", action="store_true")
    args = parser.parse_args()
    args.input = args.input.resolve()
    args.output = args.output.resolve()
    source = json.loads(args.input.read_text(encoding="utf-8"))
    records = source["papers"][: args.limit] if args.limit else source["papers"]
    completed: dict[str, dict] = {}
    if args.output.exists() and not args.refresh:
        old = json.loads(args.output.read_text(encoding="utf-8"))
        completed = {item["arxiv_id"]: item for item in old.get("papers", [])}
    pending = [item for item in records if item["arxiv_id"] not in completed]
    for offset in range(0, len(pending), args.batch_size):
        batch = pending[offset : offset + args.batch_size]
        try:
            classified = classify_batch(
                batch,
                url=args.ollama_url,
                model=args.model,
                timeout=args.timeout,
                retries=args.retries,
            )
        except Exception:
            if len(batch) == 1:
                raise
            classified = []
            for record in batch:
                classified.extend(
                    classify_batch(
                        [record],
                        url=args.ollama_url,
                        model=args.model,
                        timeout=args.timeout,
                        retries=args.retries,
                    )
                )
        for source_record, classification in zip(batch, classified, strict=True):
            completed[source_record["arxiv_id"]] = {**source_record, "analysis": classification}
        ordered = [completed[item["arxiv_id"]] for item in records if item["arxiv_id"] in completed]
        write_json(
            args.output,
            {
                "generated_at": date.today().isoformat(),
                "model": args.model,
                "source": str(args.input.relative_to(ROOT)) if args.input.is_relative_to(ROOT) else str(args.input),
                "total": len(records),
                "completed": len(ordered),
                "papers": ordered,
            },
        )
        print(f"classified {len(ordered)}/{len(records)}", flush=True)
    return 0
 if __name__ == "__main__":
    raise SystemExit(main())
@@ -0,0 +1,156 @@
 #!/usr/bin/env python3
 """Build an abstract-level research corpus for one local paper topic."""
 from __future__ import annotations
 import argparse
 import json
 import re
 import time
 import urllib.error
 import urllib.parse
 import urllib.request
 import xml.etree.ElementTree as ET
 from datetime import date
 from pathlib import Path
 ROOT = Path(__file__).resolve().parents[2]
 API_URL = "https://export.arxiv.org/api/query"
 NS = {"atom": "http://www.w3.org/2005/Atom"}
 ARXIV_ID_RE = re.compile(r"arxiv\.org/(?:abs|html|pdf)/([0-9]{4}\.[0-9]{4,5})")
 def normalize(value: str) -> str:
    return re.sub(r"\s+", " ", value).strip()
 def extract_arxiv_id(url: str) -> str | None:
    match = ARXIV_ID_RE.search(url or "")
    return match.group(1) if match else None
 def parse_feed(payload: bytes) -> dict[str, dict]:
    root = ET.fromstring(payload)
    records: dict[str, dict] = {}
    for entry in root.findall("atom:entry", NS):
        raw_id = entry.findtext("atom:id", default="", namespaces=NS)
        paper_id = raw_id.rsplit("/", 1)[-1].split("v", 1)[0]
        records[paper_id] = {
            "arxiv_id": paper_id,
            "title": normalize(entry.findtext("atom:title", default="", namespaces=NS)),
            "abstract": normalize(entry.findtext("atom:summary", default="", namespaces=NS)),
            "published": entry.findtext("atom:published", default="", namespaces=NS)[:10],
            "updated": entry.findtext("atom:updated", default="", namespaces=NS)[:10],
            "authors": [
                normalize(author.findtext("atom:name", default="", namespaces=NS))
                for author in entry.findall("atom:author", NS)
            ],
            "categories": [
                node.attrib["term"]
                for node in entry.findall("atom:category", NS)
                if node.attrib.get("term")
            ],
            "url": f"https://arxiv.org/abs/{paper_id}",
            "pdf_url": f"https://arxiv.org/pdf/{paper_id}",
        }
    return records
 def fetch_batch(ids: list[str], retries: int, timeout: int) -> dict[str, dict]:
    query = urllib.parse.urlencode({"id_list": ",".join(ids), "max_results": len(ids)})
    request = urllib.request.Request(
        f"{API_URL}?{query}",
        headers={"User-Agent": "agent-kb-research/0.1"},
    )
    for attempt in range(retries + 1):
        try:
            with urllib.request.urlopen(request, timeout=timeout) as response:
                return parse_feed(response.read())
        except (urllib.error.HTTPError, urllib.error.URLError, TimeoutError):
            if attempt >= retries:
                raise
            time.sleep(2 ** attempt)
    raise RuntimeError("unreachable retry state")
 def write_json(path: Path, payload: dict) -> None:
    path.parent.mkdir(parents=True, exist_ok=True)
    tmp = path.with_suffix(path.suffix + ".tmp")
    tmp.write_text(json.dumps(payload, ensure_ascii=False, indent=2) + "\n", encoding="utf-8")
    tmp.replace(path)
 def main() -> int:
    parser = argparse.ArgumentParser()
    parser.add_argument("--topic", default="memory")
    parser.add_argument("--index", type=Path, default=ROOT / "data" / "index.json")
    parser.add_argument("--output", type=Path)
    parser.add_argument("--batch-size", type=int, default=20)
    parser.add_argument("--sleep", type=float, default=1.0)
    parser.add_argument("--retries", type=int, default=4)
    parser.add_argument("--timeout", type=int, default=45)
    parser.add_argument("--refresh", action="store_true")
    args = parser.parse_args()
    output = args.output or ROOT / "data" / "research" / f"{args.topic}-corpus.json"
    index = json.loads(args.index.read_text(encoding="utf-8"))
    selected: dict[str, dict] = {}
    missing_ids: list[dict] = []
    for item in index:
        meta = item.get("meta") or {}
        if item.get("collection") != "papers" or args.topic not in (meta.get("topics") or []):
            continue
        paper_id = extract_arxiv_id(str(meta.get("url") or ""))
        if not paper_id:
            missing_ids.append({"path": item.get("path"), "title": item.get("title")})
            continue
        selected[paper_id] = {
            "path": item.get("path"),
            "local_title": item.get("title"),
            "local_status": meta.get("status"),
            "local_topics": meta.get("topics") or [],
            "collection_queries": meta.get("collection_queries"),
            "collection_score": meta.get("collection_score"),
        }
    cached: dict[str, dict] = {}
    if output.exists() and not args.refresh:
        old = json.loads(output.read_text(encoding="utf-8"))
        cached = {record["arxiv_id"]: record for record in old.get("papers", [])}
    fetched: dict[str, dict] = {}
    ids_to_fetch = [paper_id for paper_id in selected if paper_id not in cached]
    for offset in range(0, len(ids_to_fetch), args.batch_size):
        batch = ids_to_fetch[offset : offset + args.batch_size]
        fetched.update(fetch_batch(batch, retries=args.retries, timeout=args.timeout))
        print(f"fetched {min(offset + len(batch), len(ids_to_fetch))}/{len(ids_to_fetch)}")
        if offset + args.batch_size < len(ids_to_fetch):
            time.sleep(args.sleep)
    papers = []
    unresolved = []
    for paper_id, local in selected.items():
        remote = fetched.get(paper_id) or cached.get(paper_id)
        if not remote:
            unresolved.append({"arxiv_id": paper_id, **local})
            continue
        papers.append({**remote, **local})
    papers.sort(key=lambda item: (item.get("published") or "", item["arxiv_id"]))
    payload = {
        "topic": args.topic,
        "generated_at": date.today().isoformat(),
        "selected_from_index": len(selected),
        "resolved": len(papers),
        "missing_arxiv_id": missing_ids,
        "unresolved": unresolved,
        "papers": papers,
    }
    write_json(output, payload)
    print(json.dumps({key: payload[key] for key in ("selected_from_index", "resolved")}, indent=2))
    return 0 if not unresolved else 2
 if __name__ == "__main__":
    raise SystemExit(main())
@@ -0,0 +1,281 @@
 #!/usr/bin/env python3
 """Fast relevance screening for a broad Agent-memory research corpus."""
 from __future__ import annotations
 import argparse
 import json
 import re
 import time
 import urllib.error
 import urllib.request
 from datetime import date
 from pathlib import Path
 ROOT = Path(__file__).resolve().parents[2]
 PROMPT = """Screen papers for a research review about memory in LLM-based agents.
 Use title and abstract only. Return JSON with a `papers` array in input order.
 For each paper return:
 - arxiv_id: exact input id
 - relevance: core if memory is the main research object; supporting if memory is a substantial
  mechanism, evaluation dimension, or failure mode; peripheral if memory is incidental, means
  GPU/KV memory only, refers to human memory only, or the paper is mainly an unrelated application
 - problem: choose one: access, representation, update_forgetting, execution_state, experience_learning,
  evaluation, security_privacy, shared_memory, systems_cost, theory, application, unrelated
 - reason: at most 12 words
 Be conservative and output JSON only."""
 RELEVANCE_VALUES = {"core", "supporting", "peripheral"}
 PROBLEM_VALUES = {
    "access",
    "representation",
    "update_forgetting",
    "execution_state",
    "experience_learning",
    "evaluation",
    "security_privacy",
    "shared_memory",
    "systems_cost",
    "theory",
    "application",
    "unrelated",
 }
 PROBLEM_ALIASES = {
    "adaptation": "experience_learning",
    "continual_learning": "experience_learning",
    "episodic": "representation",
    "episodic_memory": "representation",
    "experience": "experience_learning",
    "forgetting": "update_forgetting",
    "implicit_personalization": "application",
    "memory": "representation",
    "privacy": "security_privacy",
    "security": "security_privacy",
 }
 def write_json(path: Path, payload: dict) -> None:
    path.parent.mkdir(parents=True, exist_ok=True)
    tmp = path.with_suffix(path.suffix + ".tmp")
    tmp.write_text(json.dumps(payload, ensure_ascii=False, indent=2) + "\n", encoding="utf-8")
    tmp.replace(path)
 def generate(url: str, model: str, records: list[dict], timeout: int, retries: int) -> list[dict]:
    compact = [
        {"arxiv_id": item["arxiv_id"], "title": item["title"], "abstract": item["abstract"]}
        for item in records
    ]
    body = {
        "model": model,
        "prompt": f"{PROMPT}\n\nINPUT:\n{json.dumps(compact, ensure_ascii=False)}",
        "stream": False,
        "format": "json",
        "options": {"temperature": 0, "num_predict": 4000},
    }
    request = urllib.request.Request(
        f"{url.rstrip('/')}/api/generate",
        data=json.dumps(body).encode("utf-8"),
        headers={"Content-Type": "application/json"},
        method="POST",
    )
    for attempt in range(retries + 1):
        try:
            with urllib.request.urlopen(request, timeout=timeout) as response:
                result = json.loads(response.read().decode("utf-8"))
            decoded = json.loads(result["response"])
            if isinstance(decoded, list):
                parsed = decoded
            elif isinstance(decoded, dict) and isinstance(decoded.get("papers"), list):
                parsed = decoded["papers"]
            elif isinstance(decoded, dict) and isinstance(decoded.get("results"), list):
                parsed = decoded["results"]
            elif isinstance(decoded, dict) and decoded.get("arxiv_id"):
                parsed = [decoded]
            else:
                raise ValueError("response lacks a recognized result array")
            expected = [item["arxiv_id"] for item in records]
            by_id = {str(item.get("arxiv_id")): item for item in parsed}
            if any(paper_id not in by_id for paper_id in expected):
                raise ValueError("response ids do not match input")
            return [{**by_id[paper_id], "source": "model"} for paper_id in expected]
        except (
            urllib.error.HTTPError,
            urllib.error.URLError,
            TimeoutError,
            json.JSONDecodeError,
            KeyError,
            ValueError,
        ):
            if attempt >= retries:
                raise
            time.sleep(2 ** attempt)
    raise RuntimeError("unreachable retry state")
 def heuristic_screen(record: dict) -> dict:
    title = record["title"].lower()
    text = f"{record['title']} {record['abstract']}".lower()
    title_memory = bool(re.search(r"memory|memori|remember|forget|stateful", title))
    agent_memory = bool(
        re.search(r"agent(?:ic)? memory|memory (?:for|in|of) .*agent|long-term .*agent", text)
    )
    hardware_only = bool(re.search(r"gpu memory|memory footprint|memory bandwidth|memory efficient", title))
    if title_memory and not hardware_only:
        relevance = "core"
    elif agent_memory:
        relevance = "supporting"
    else:
        relevance = "peripheral"
    if re.search(r"attack|poison|privacy|security|trust|provenance|leak", text):
        problem = "security_privacy"
    elif re.search(r"benchmark|evaluation|diagnos|measure", title):
        problem = "evaluation"
    elif re.search(r"update|forget|consolidat|stale|supersed|write", text):
        problem = "update_forgetting"
    elif re.search(r"skill|experience|reflect|self-evolv|continual learn", text):
        problem = "experience_learning"
    elif re.search(r"execution state|task state|workflow state|long-horizon", text):
        problem = "execution_state"
    elif re.search(r"graph|hierarch|represent|structure|database|document", text):
        problem = "representation"
    elif re.search(r"retriev|search|recall|access", text):
        problem = "access"
    elif re.search(r"multi-agent|shared memory|collective memory", text):
        problem = "shared_memory"
    elif re.search(r"cost|latency|token|efficient|compress", text):
        problem = "systems_cost"
    else:
        problem = "application" if relevance != "peripheral" else "unrelated"
    return {
        "arxiv_id": record["arxiv_id"],
        "relevance": relevance,
        "problem": problem,
        "reason": "deterministic fallback after model-format failure",
        "source": "heuristic",
    }
 def normalize_analysis(record: dict, analysis: dict | None) -> dict:
    """Constrain model output to the documented screening schema."""
    fallback = heuristic_screen(record)
    analysis = analysis if isinstance(analysis, dict) else {}
    relevance = str(analysis.get("relevance", "")).strip().lower()
    if relevance not in RELEVANCE_VALUES:
        relevance = fallback["relevance"]
    problem = str(analysis.get("problem", "")).strip().lower()
    problem = PROBLEM_ALIASES.get(problem, problem)
    if problem not in PROBLEM_VALUES:
        problem = fallback["problem"]
    reason = str(analysis.get("reason", "")).strip() or fallback["reason"]
    source = str(analysis.get("source", "")).strip().lower()
    if source not in {"model", "heuristic"}:
        source = "model" if analysis else "heuristic"
    return {
        "arxiv_id": record["arxiv_id"],
        "relevance": relevance,
        "problem": problem,
        "reason": reason,
        "source": source,
    }
 def main() -> int:
    parser = argparse.ArgumentParser()
    parser.add_argument(
        "--input",
        type=Path,
        default=ROOT / "data" / "research" / "agent-memory-expanded.json",
    )
    parser.add_argument(
        "--output",
        type=Path,
        default=ROOT / "data" / "research" / "agent-memory-screened.json",
    )
    parser.add_argument("--ollama-url", default="http://192.168.1.10:11434")
    parser.add_argument("--model", default="ChatGPT-5.6:Luna")
    parser.add_argument("--batch-size", type=int, default=24)
    parser.add_argument("--timeout", type=int, default=180)
    parser.add_argument("--retries", type=int, default=2)
    parser.add_argument("--refresh", action="store_true")
    args = parser.parse_args()
    args.input = args.input.resolve()
    args.output = args.output.resolve()
    source = json.loads(args.input.read_text(encoding="utf-8"))
    records = source["papers"]
    completed: dict[str, dict] = {}
    if args.output.exists() and not args.refresh:
        previous = json.loads(args.output.read_text(encoding="utf-8"))
        completed = {
            item["arxiv_id"]: {
                **item,
                "screen": normalize_analysis(item, item.get("screen")),
            }
            for item in previous.get("papers", [])
        }
    pending = [item for item in records if item["arxiv_id"] not in completed]
    for offset in range(0, len(pending), args.batch_size):
        batch = pending[offset : offset + args.batch_size]
        try:
            screened = generate(
                args.ollama_url, args.model, batch, timeout=args.timeout, retries=args.retries
            )
        except Exception:
            screened = []
            for record in batch:
                try:
                    screened.extend(
                        generate(
                            args.ollama_url,
                            args.model,
                            [record],
                            timeout=args.timeout,
                            retries=args.retries,
                        )
                    )
                except Exception:
                    screened.append(heuristic_screen(record))
        for record, analysis in zip(batch, screened, strict=True):
            completed[record["arxiv_id"]] = {
                **record,
                "screen": normalize_analysis(record, analysis),
            }
        ordered = [completed[item["arxiv_id"]] for item in records if item["arxiv_id"] in completed]
        write_json(
            args.output,
            {
                "generated_at": date.today().isoformat(),
                "model": args.model,
                "source": str(args.input.relative_to(ROOT)),
                "total": len(records),
                "completed": len(ordered),
                "papers": ordered,
            },
        )
        print(f"screened {len(ordered)}/{len(records)}", flush=True)
    # Rewrites an already complete file after schema normalization as well.
    ordered = [completed[item["arxiv_id"]] for item in records if item["arxiv_id"] in completed]
    write_json(
        args.output,
        {
            "generated_at": date.today().isoformat(),
            "model": args.model,
            "source": str(args.input.relative_to(ROOT)),
            "total": len(records),
            "completed": len(ordered),
            "papers": ordered,
        },
    )
    return 0
 if __name__ == "__main__":
    raise SystemExit(main())
@@ -0,0 +1,194 @@
 #!/usr/bin/env python3
 """Collect a paginated, abstract-level Agent-memory research corpus from arXiv."""
 from __future__ import annotations
 import argparse
 import json
 import re
 import time
 import urllib.error
 import urllib.parse
 import urllib.request
 import xml.etree.ElementTree as ET
 from collections import defaultdict
 from datetime import date
 from pathlib import Path
 ROOT = Path(__file__).resolve().parents[2]
 API_URL = "https://export.arxiv.org/api/query"
 NS = {
    "atom": "http://www.w3.org/2005/Atom",
    "opensearch": "http://a9.com/-/spec/opensearch/1.1/",
 }
 QUERIES = [
    ("agent-memory", 'all:"agent memory" OR all:"memory agent" OR all:"memory agents"'),
    (
        "long-term-memory",
        '(all:"long-term memory" OR all:"persistent memory") AND '
        '(all:"LLM agent" OR all:"language agent" OR all:"large language model agent")',
    ),
    (
        "episodic-memory",
        'all:"episodic memory" AND (all:"LLM" OR all:"large language model") AND all:"agent"',
    ),
    (
        "procedural-memory",
        '(all:"procedural memory" OR all:"skill memory") AND '
        '(all:"LLM agent" OR all:"language agent")',
    ),
    (
        "memory-evaluation",
        '(all:"memory benchmark" OR all:"memory evaluation") AND '
        '(all:"LLM agent" OR all:"language agent" OR all:"agent memory")',
    ),
 ]
 def normalize(value: str) -> str:
    return re.sub(r"\s+", " ", value).strip()
 def arxiv_id(raw: str) -> str:
    return raw.rsplit("/", 1)[-1].split("v", 1)[0]
 def date_query(query: str, start: str, end: str) -> str:
    return (
        f"({query}) AND submittedDate:[{start.replace('-', '')}0000 "
        f"TO {end.replace('-', '')}2359]"
    )
 def fetch_page(
    query: str,
    start_date: str,
    end_date: str,
    offset: int,
    page_size: int,
    retries: int,
    timeout: int,
 ) -> bytes:
    params = {
        "search_query": date_query(query, start_date, end_date),
        "start": offset,
        "max_results": page_size,
        "sortBy": "submittedDate",
        "sortOrder": "ascending",
    }
    request = urllib.request.Request(
        f"{API_URL}?{urllib.parse.urlencode(params)}",
        headers={"User-Agent": "agent-kb-research/0.1"},
    )
    for attempt in range(retries + 1):
        try:
            with urllib.request.urlopen(request, timeout=timeout) as response:
                return response.read()
        except (urllib.error.HTTPError, urllib.error.URLError, TimeoutError):
            if attempt >= retries:
                raise
            time.sleep(2 ** attempt)
    raise RuntimeError("unreachable retry state")
 def parse_page(payload: bytes) -> tuple[int, list[dict]]:
    root = ET.fromstring(payload)
    total = int(root.findtext("opensearch:totalResults", default="0", namespaces=NS))
    records = []
    for entry in root.findall("atom:entry", NS):
        paper_id = arxiv_id(entry.findtext("atom:id", default="", namespaces=NS))
        records.append(
            {
                "arxiv_id": paper_id,
                "title": normalize(entry.findtext("atom:title", default="", namespaces=NS)),
                "abstract": normalize(entry.findtext("atom:summary", default="", namespaces=NS)),
                "published": entry.findtext("atom:published", default="", namespaces=NS)[:10],
                "updated": entry.findtext("atom:updated", default="", namespaces=NS)[:10],
                "authors": [
                    normalize(author.findtext("atom:name", default="", namespaces=NS))
                    for author in entry.findall("atom:author", NS)
                ],
                "categories": [
                    node.attrib["term"]
                    for node in entry.findall("atom:category", NS)
                    if node.attrib.get("term")
                ],
                "url": f"https://arxiv.org/abs/{paper_id}",
                "pdf_url": f"https://arxiv.org/pdf/{paper_id}",
            }
        )
    return total, records
 def write_json(path: Path, payload: dict) -> None:
    path.parent.mkdir(parents=True, exist_ok=True)
    tmp = path.with_suffix(path.suffix + ".tmp")
    tmp.write_text(json.dumps(payload, ensure_ascii=False, indent=2) + "\n", encoding="utf-8")
    tmp.replace(path)
 def main() -> int:
    parser = argparse.ArgumentParser()
    parser.add_argument("--from-date", default="2025-01-01")
    parser.add_argument("--to-date", default=date.today().isoformat())
    parser.add_argument("--page-size", type=int, default=100)
    parser.add_argument("--max-per-query", type=int, default=2000)
    parser.add_argument("--sleep", type=float, default=1.0)
    parser.add_argument("--retries", type=int, default=4)
    parser.add_argument("--timeout", type=int, default=45)
    parser.add_argument(
        "--output",
        type=Path,
        default=ROOT / "data" / "research" / "agent-memory-expanded.json",
    )
    args = parser.parse_args()
    records: dict[str, dict] = {}
    matched_queries: defaultdict[str, set[str]] = defaultdict(set)
    query_totals = {}
    for name, query in QUERIES:
        offset = 0
        while offset < args.max_per_query:
            payload = fetch_page(
                query,
                args.from_date,
                args.to_date,
                offset,
                args.page_size,
                args.retries,
                args.timeout,
            )
            total, page = parse_page(payload)
            query_totals[name] = total
            for record in page:
                records[record["arxiv_id"]] = record
                matched_queries[record["arxiv_id"]].add(name)
            offset += len(page)
            print(f"{name}: {min(offset, total)}/{total}", flush=True)
            if not page or offset >= total:
                break
            time.sleep(args.sleep)
        time.sleep(args.sleep)
    papers = [
        {**record, "matched_queries": sorted(matched_queries[paper_id])}
        for paper_id, record in records.items()
    ]
    papers.sort(key=lambda item: (item["published"], item["arxiv_id"]))
    output = {
        "generated_at": date.today().isoformat(),
        "from_date": args.from_date,
        "to_date": args.to_date,
        "query_totals": query_totals,
        "unique_papers": len(papers),
        "papers": papers,
    }
    write_json(args.output, output)
    print(json.dumps({"query_totals": query_totals, "unique_papers": len(papers)}, indent=2))
    return 0
 if __name__ == "__main__":
    raise SystemExit(main())