Add Agent Memory learning pilot

2026-07-10 18:01:44 +08:00
parent 15b5cec933
commit 36c631e026
22 changed files with 1195 additions and 224 deletions
@@ -12,47 +12,49 @@ source: arxiv
 collected_at: 2026-07-08
 published_at: 2026-03-12
 updated_at: 2026-03-12
-status: queued
+status: skimmed
 relevance: high
 topics:
  - memory
  - rag
 methods:
-  -
+  - verified-experience-distillation
+  - instruction-rewriting
+  - promotion-gate
 benchmarks:
-  -
+  - CUAD
+  - HotpotQA
 models:
  -
 datasets:
  - cs.AI
 related_concepts:
-  -
+  - procedural-memory
+  - prompt-compilation
 related_jobs:
  -
 related_experiments:
-  -
+  - KC-001-agent-memory-pilot
 related_projects:
-  -
+  - learning/agent-memory
 collection_score: 13
 collection_queries: language-agent
 ---

 ## One-line Takeaway

-Auto-collected from arXiv because it matched the Agent collection queries. Needs human skim.
+对稳定重复任务，经过验证的经验可以被编译成行为规则，而不必每次作为 retrieved context 注入。

-## Why Collected
+## Pilot Skim

- matched queries: language-agent
- inferred topics: memory, rag
- arXiv categories: cs.AI
- collection score: 13
+- problem: 有更多历史不代表 Agent 知道哪些经验应改变未来行为。
+- method: Atlas 用多层 promotion 和 verification 把经验写成 system prompt 子规则。
+- evidence: CUAD token F1 +8.7pp，HotpotQA joint F1 +3.16pp；跨模型复用同一 prompt 时 +2.31pp。
+- boundary: evolved prompt 更长且缺少 length-matched control；改进严格受训练信号覆盖范围限制。

-## Review Checklist
+## Used In

- Does this paper directly inform Agent architecture, evaluation, memory, tools, safety, coding agents, GUI/browser agents, or multi-agent workflows?
- Does it include a benchmark, dataset, code, or reproducible experimental setup?
- Should it be promoted from `queued` to `skimmed` or `summarized`?
+- [Agent Memory evidence matrix](../../learning/agent-memory/evidence-matrix.md)

 ## Links

@@ -12,7 +12,7 @@ source: arxiv
 collected_at: 2026-07-08
 published_at: 2026-05-08
 updated_at: 2026-05-14
-status: queued
+status: skimmed
 relevance: high
 topics:
  - agent-evaluation
@@ -23,41 +23,47 @@ topics:
  - reasoning
  - tool-use
 methods:
-  -
+  - provenance-dag
+  - td-lambda
+  - q-guided-retrieval
 benchmarks:
-  -
+  - LLAB
+  - LiveCodeBench
+  - MMMU-Pro
+  - ERQA
+  - GPQA-Diamond
+  - BFCL
 models:
  -
 datasets:
  - cs.AI
 related_concepts:
-  -
+  - memory-credit-assignment
+  - procedural-memory
 related_jobs:
  -
 related_experiments:
-  -
+  - KC-001-agent-memory-pilot
 related_projects:
-  -
+  - learning/agent-memory
 collection_score: 20
 collection_queries: function-calling
 ---

 ## One-line Takeaway

-Auto-collected from arXiv because it matched the Agent collection queries. Needs human skim.
+Memory 的长期价值应沿“哪些旧记忆帮助生成了新记忆”的 provenance DAG 反向归因，而不只奖励最后一次检索。

-## Why Collected
+## Pilot Skim

- matched queries: function-calling
- inferred topics: agent-evaluation, computer-use, embodied-agent, memory, rag, reasoning, tool-use
- arXiv categories: cs.AI
- collection score: 20
+- problem: 单步 memory value update 忽略跨多次任务形成的经验依赖链。
+- method: 用 TD(lambda) 沿 provenance DAG 回传 credit，结构深度代替时间步距离。
+- evidence: 六个 benchmark 上领先或并列；相对 MemRL 的收益在深链任务最高达到 5.7pp，单步任务最低约 0.77pp。
+- boundary: 假设 memory 单调增长，未解决 consolidation/deletion；DAG 存储和 BFS 更新有持续成本。

-## Review Checklist
+## Used In

- Does this paper directly inform Agent architecture, evaluation, memory, tools, safety, coding agents, GUI/browser agents, or multi-agent workflows?
- Does it include a benchmark, dataset, code, or reproducible experimental setup?
- Should it be promoted from `queued` to `skimmed` or `summarized`?
+- [Agent Memory evidence matrix](../../learning/agent-memory/evidence-matrix.md)

 ## Links

@@ -12,48 +12,50 @@ source: arxiv
 collected_at: 2026-07-08
 published_at: 2026-05-11
 updated_at: 2026-05-11
-status: queued
+status: skimmed
 relevance: high
 topics:
  - agent-evaluation
  - memory
  - planning
 methods:
-  -
+  - decision-rate-distortion
+  - k-slot-memory
+  - certified-memory-split
 benchmarks:
-  -
+  - LoCoMo
+  - synthetic-decision-diagnostics
 models:
  -
 datasets:
  - cs.AI
 related_concepts:
-  -
+  - decision-sufficient-state
+  - selective-forgetting
 related_jobs:
  -
 related_experiments:
-  -
+  - KC-001-agent-memory-pilot
 related_projects:
-  -
+  - learning/agent-memory
 collection_score: 13
 collection_queries: language-agent
 ---

 ## One-line Takeaway

-Auto-collected from arXiv because it matched the Agent collection queries. Needs human skim.
+Memory 应保留会改变决策的历史差异，而不是优先保留描述最相似或最完整的记录。

-## Why Collected
+## Pilot Skim

- matched queries: language-agent
- inferred topics: agent-evaluation, memory, planning
- arXiv categories: cs.AI
- collection score: 13
+- problem: relevance、salience 和 summary fidelity 没有直接刻画 memory 对未来决策的价值。
+- method: 在 K 个 runtime state 预算下定义 decision distortion、forgetting boundary，并按 certified conflict 拆分 state。
+- evidence: LoCoMo 子集上描述相似度与 evidence compatibility 的相关仅 0.103；matched budget 下 DeMem gold evidence recall 为 83%，描述检索为 66%。
+- boundary: 理论主要在 contextual bandit 和有限状态设定中建立，实际 LLM 冲突判定仍依赖 judge 与近似算法。

-## Review Checklist
+## Used In

- Does this paper directly inform Agent architecture, evaluation, memory, tools, safety, coding agents, GUI/browser agents, or multi-agent workflows?
- Does it include a benchmark, dataset, code, or reproducible experimental setup?
- Should it be promoted from `queued` to `skimmed` or `summarized`?
+- [Agent Memory evidence matrix](../../learning/agent-memory/evidence-matrix.md)

 ## Links

@@ -12,7 +12,7 @@ source: arxiv
 collected_at: 2026-07-08
 published_at: 2026-05-14
 updated_at: 2026-05-14
-status: queued
+status: skimmed
 relevance: high
 topics:
  - agent-evaluation
@@ -20,42 +20,44 @@ topics:
  - memory
  - tool-use
 methods:
-  -
+  - cryptographic-provenance
+  - derivation-lineage
+  - sensitive-action-gate
 benchmarks:
-  -
+  - custom-cross-session-memory-harness
+  - AgentDojo-bridge
 models:
  -
 datasets:
  - cs.CR
  - cs.AI
 related_concepts:
-  -
+  - memory-provenance
+  - memory-poisoning
 related_jobs:
  -
 related_experiments:
-  -
+  - KC-001-agent-memory-pilot
 related_projects:
-  -
+  - learning/agent-memory
 collection_score: 17
 collection_queries: agent-memory
 ---

 ## One-line Takeaway

-Auto-collected from arXiv because it matched the Agent collection queries. Needs human skim.
+签名只能证明谁写入；敏感行动还必须检查当前理由是否沿派生链继承了不可信来源。

-## Why Collected
+## Pilot Skim

- matched queries: agent-memory
- inferred topics: agent-evaluation, computer-use, memory, tool-use
- arXiv categories: cs.CR, cs.AI
- collection score: 17
+- problem: 恶意外部内容可以经多次 memory derivation 被语义洗白，未来再触发敏感动作。
+- method: 为条目附加签名与 derivation DAG，把来源 trust 传播到 action gate。
+- evidence: 自定义 deterministic harness 的三类攻击中，只有 lineage 方案全部达到 0 ASR。
+- boundary: headline 是机制隔离 benchmark，不是完整公共 benchmark；未完成大规模 adaptive attack，且假设模型推理路径可信。

-## Review Checklist
+## Used In

- Does this paper directly inform Agent architecture, evaluation, memory, tools, safety, coding agents, GUI/browser agents, or multi-agent workflows?
- Does it include a benchmark, dataset, code, or reproducible experimental setup?
- Should it be promoted from `queued` to `skimmed` or `summarized`?
+- [Agent Memory evidence matrix](../../learning/agent-memory/evidence-matrix.md)

 ## Links

@@ -12,7 +12,7 @@ source: arxiv
 collected_at: 2026-07-08
 published_at: 2026-06-03
 updated_at: 2026-06-03
-status: queued
+status: skimmed
 relevance: high
 topics:
  - agent-evaluation
@@ -21,41 +21,45 @@ topics:
  - rag
  - tool-use
 methods:
-  -
+  - agentic-memory-harness
+  - schema-diagnostics
+  - active-retrieval
 benchmarks:
-  -
+  - LoCoMo
+  - HotpotQA
+  - AMABench
+  - ALFWorld
 models:
  -
 datasets:
  - cs.AI
 related_concepts:
-  -
+  - schema-commitment
+  - agentic-retrieval
 related_jobs:
  -
 related_experiments:
-  -
+  - KC-001-agent-memory-pilot
 related_projects:
-  -
+  - learning/agent-memory
 collection_score: 19
 collection_queries: agent-memory
 ---

 ## One-line Takeaway

-Auto-collected from arXiv because it matched the Agent collection queries. Needs human skim.
+复杂索引会在写入时承诺 schema；保留原始证据并让 Agent 按问题主动读取，往往有更好的跨场景通用性。

-## Why Collected
+## Pilot Skim

- matched queries: agent-memory
- inferred topics: agent-evaluation, memory, planning, rag, tool-use
- arXiv categories: cs.AI
- collection score: 19
+- problem: 为单一场景设计的 memory 在对话、QA、trajectory 和动态任务之间难以泛化。
+- method: 诊断 representation loss 和 retrieval loss，并提出带主动工具循环的 AutoMEM。
+- evidence: AutoMEM 在 LoCoMo 为 67.3，对 long context 的 61.5；ALFWorld 诊断显示 golden procedure 仍低于训练后的 actor。
+- boundary: agentic harness 不是每个场景最省 token；动态任务上限可能在 policy 而非 memory。

-## Review Checklist
+## Used In

- Does this paper directly inform Agent architecture, evaluation, memory, tools, safety, coding agents, GUI/browser agents, or multi-agent workflows?
- Does it include a benchmark, dataset, code, or reproducible experimental setup?
- Should it be promoted from `queued` to `skimmed` or `summarized`?
+- [Agent Memory evidence matrix](../../learning/agent-memory/evidence-matrix.md)

 ## Links

@@ -12,7 +12,7 @@ source: arxiv
 collected_at: 2026-07-08
 published_at: 2026-06-04
 updated_at: 2026-06-04
-status: queued
+status: skimmed
 relevance: high
 topics:
  - computer-use
@@ -21,41 +21,42 @@ topics:
  - rag
  - tool-use
 methods:
-  -
+  - hierarchical-execution-state-tree
+  - grow-compress-maintain-revise
+  - error-branch-isolation
 benchmarks:
-  -
+  - MemoryArena
 models:
  -
 datasets:
  - cs.AI
 related_concepts:
-  -
+  - execution-state
+  - long-horizon-agent
 related_jobs:
  -
 related_experiments:
-  -
+  - KC-001-agent-memory-pilot
 related_projects:
-  -
+  - learning/agent-memory
 collection_score: 16
 collection_queries: agent-memory, rag-agent
 ---

 ## One-line Takeaway

-Auto-collected from arXiv because it matched the Agent collection queries. Needs human skim.
+长程任务需要维护 root-to-current 执行状态并隔离错误分支，而不是从相似记录中临时拼装上下文。

-## Why Collected
+## Pilot Skim

- matched queries: agent-memory, rag-agent
- inferred topics: computer-use, memory, planning, rag, tool-use
- arXiv categories: cs.AI
- collection score: 16
+- problem: 相似度组织会割裂执行依赖，并把有效和错误轨迹重新混入同一上下文。
+- method: 两层 state tree 保存 action-observation 与 subgoal summary，通过 Grow、Compress、Maintain、Revise 管理当前路径。
+- evidence: MemoryArena 上相对 baselines 平均提高 7.8 至 20.4pp，并比长上下文减少 55.1% token。
+- boundary: 证据针对 interdependent long-horizon tasks，不直接代表事实问答或个性化 memory。

-## Review Checklist
+## Used In

- Does this paper directly inform Agent architecture, evaluation, memory, tools, safety, coding agents, GUI/browser agents, or multi-agent workflows?
- Does it include a benchmark, dataset, code, or reproducible experimental setup?
- Should it be promoted from `queued` to `skimmed` or `summarized`?
+- [Agent Memory evidence matrix](../../learning/agent-memory/evidence-matrix.md)

 ## Links

@@ -12,7 +12,7 @@ source: arxiv
 collected_at: 2026-07-08
 published_at: 2026-06-09
 updated_at: 2026-06-09
-status: queued
+status: skimmed
 relevance: high
 topics:
  - agent-evaluation
@@ -20,42 +20,44 @@ topics:
  - rag
  - tool-use
 methods:
-  -
+  - topic-documents
+  - buffered-consolidation
+  - agentic-retrieval
 benchmarks:
-  -
+  - MemoryAgentBench
+  - LongMemEval
 models:
  -
 datasets:
  - cs.AI
  - cs.CL
 related_concepts:
-  -
+  - semantic-memory
+  - selective-forgetting
 related_jobs:
  -
 related_experiments:
-  -
+  - KC-001-agent-memory-pilot
 related_projects:
-  -
+  - learning/agent-memory
 collection_score: 18
 collection_queries: agent-memory
 ---

 ## One-line Takeaway

-Auto-collected from arXiv because it matched the Agent collection queries. Needs human skim.
+长期事实应被维护为带证据和时间的主题文档，并允许 Agent 多步检查，而不是只保存孤立记录和一次检索结果。

-## Why Collected
+## Pilot Skim

- matched queries: agent-memory
- inferred topics: agent-evaluation, memory, rag, tool-use
- arXiv categories: cs.AI, cs.CL
- collection score: 18
+- problem: 孤立记录难以聚合证据、修订事实和维护跨 session 一致性。
+- method: observation 先进入 buffer，再巩固到 topic documents；读取时迭代 search、grep 和 read_lines。
+- evidence: MemoryAgentBench overall 64.7；LongMemEval 上 summary-only 41.7、summary+BM25 76.0、agentic 79.3。
+- boundary: multi-hop selective forgetting 仍困难；多步 retrieval 和周期维护增加计算成本。

-## Review Checklist
+## Used In

- Does this paper directly inform Agent architecture, evaluation, memory, tools, safety, coding agents, GUI/browser agents, or multi-agent workflows?
- Does it include a benchmark, dataset, code, or reproducible experimental setup?
- Should it be promoted from `queued` to `skimmed` or `summarized`?
+- [Agent Memory evidence matrix](../../learning/agent-memory/evidence-matrix.md)

 ## Links

@@ -12,7 +12,7 @@ source: arxiv
 collected_at: 2026-07-08
 published_at: 2026-06-10
 updated_at: 2026-06-10
-status: queued
+status: skimmed
 relevance: high
 topics:
  - agent-evaluation
@@ -23,9 +23,13 @@ topics:
  - rag
  - reasoning
 methods:
-  -
+  - hierarchical-memory-workspace
+  - memory-skill-evolution
+  - rl-retrieval-agent
 benchmarks:
-  -
+  - ALFWorld
+  - LoCoMo
+  - LongMemEval
 models:
  -
 datasets:
@@ -33,33 +37,32 @@ datasets:
  - cs.CL
  - cs.LG
 related_concepts:
-  -
+  - memory-organization
+  - agentic-retrieval
 related_jobs:
  -
 related_experiments:
-  -
+  - KC-001-agent-memory-pilot
 related_projects:
-  -
+  - learning/agent-memory
 collection_score: 18
 collection_queries: agent-memory
 ---

 ## One-line Takeaway

-Auto-collected from arXiv because it matched the Agent collection queries. Needs human skim.
+Memory construction 和 retrieval 应分开优化：强模型负责长期组织，轻量策略负责沿层级 workspace 取回最小充分证据。

-## Why Collected
+## Pilot Skim

- matched queries: agent-memory
- inferred topics: agent-evaluation, computer-use, embodied-agent, memory, planning, rag, reasoning
- arXiv categories: cs.AI, cs.CL, cs.LG
- collection score: 18
+- problem: 联合优化构建与检索会让稀疏任务 reward 无法归因，也混淆两个不同时间尺度的职责。
+- method: 高层 manager 组织带 provenance 的文件层级，retriever 用 Bash-like tools 导航并通过 RL 优化。
+- evidence: ALFWorld strict context 下达到 56.7/73.9 SR；小型 Qwen retriever 在对话数据训练后迁移到其他域。
+- boundary: weak manager 造成的组织缺陷无法由强 retriever 补回，系统依赖高能力 construction model。

-## Review Checklist
+## Used In

- Does this paper directly inform Agent architecture, evaluation, memory, tools, safety, coding agents, GUI/browser agents, or multi-agent workflows?
- Does it include a benchmark, dataset, code, or reproducible experimental setup?
- Should it be promoted from `queued` to `skimmed` or `summarized`?
+- [Agent Memory evidence matrix](../../learning/agent-memory/evidence-matrix.md)

 ## Links

@@ -12,7 +12,7 @@ source: arxiv
 collected_at: 2026-07-08
 published_at: 2026-06-12
 updated_at: 2026-06-12
-status: queued
+status: skimmed
 relevance: high
 topics:
  - agent-evaluation
@@ -21,41 +21,42 @@ topics:
  - reasoning
  - workflow-agent
 methods:
-  -
+  - budget-matched-baseline
+  - online-augmentation-audit
 benchmarks:
-  -
+  - WebArena
+  - WorkArena-L1
 models:
  -
 datasets:
  - cs.CL
 related_concepts:
-  -
+  - memory-cost-accounting
+  - evaluation-hygiene
 related_jobs:
  -
 related_experiments:
-  -
+  - KC-001-agent-memory-pilot
 related_projects:
-  -
+  - learning/agent-memory
 collection_score: 14
 collection_queries: web-gui-agent
 ---

 ## One-line Takeaway

-Auto-collected from arXiv because it matched the Agent collection queries. Needs human skim.
+在线 memory 或 skill 的收益必须与“把相同推理预算直接交给基础 Agent”比较，不能只对比相同步数。

-## Why Collected
+## Pilot Skim

- matched queries: web-gui-agent
- inferred topics: agent-evaluation, computer-use, memory, reasoning, workflow-agent
- arXiv categories: cs.CL
- collection score: 14
+- problem: augmentation 论文常忽略 induction、retrieval、verification 和 prompt injection 的额外 token。
+- method: 用 Vanilla-IB 把近似相同预算用于更多 observe-act steps，并报告多 run 方差。
+- evidence: WebArena 三域三模型中，Vanilla-IB 平均 SR 均最高且通常 token 更少；WorkArena-L1 上仍处于 success-cost frontier。
+- boundary: 研究对象是 online augmentation；可离线摊销的 skill/memory 需要另一种成本模型。

-## Review Checklist
+## Used In

- Does this paper directly inform Agent architecture, evaluation, memory, tools, safety, coding agents, GUI/browser agents, or multi-agent workflows?
- Does it include a benchmark, dataset, code, or reproducible experimental setup?
- Should it be promoted from `queued` to `skimmed` or `summarized`?
+- [Agent Memory evidence matrix](../../learning/agent-memory/evidence-matrix.md)

 ## Links

@@ -12,7 +12,7 @@ source: arxiv
 collected_at: 2026-07-08
 published_at: 2026-06-23
 updated_at: 2026-06-23
-status: queued
+status: skimmed
 relevance: high
 topics:
  - agent-evaluation
@@ -22,41 +22,44 @@ topics:
  - reasoning
  - tool-use
 methods:
-  -
+  - memory-transition-verification
+  - transition-ranked-grpo
+  - write-revise-prune
 benchmarks:
-  -
+  - MemoryAgentBench
+  - HaluMem
+  - Mem-alpha
 models:
  -
 datasets:
  - cs.AI
 related_concepts:
-  -
+  - memory-consolidation
+  - transition-verification
 related_jobs:
  -
 related_experiments:
-  -
+  - KC-001-agent-memory-pilot
 related_projects:
-  -
+  - learning/agent-memory
 collection_score: 19
 collection_queries: agent-memory
 ---

 ## One-line Takeaway

-Auto-collected from arXiv because it matched the Agent collection queries. Needs human skim.
+持久 memory 需要在每次 WRITE、REVISE、PRUNE 后检查 coverage、preservation 和 faithfulness，不能只看最终回答是否正确。

-## Why Collected
+## Pilot Skim

- matched queries: agent-memory
- inferred topics: agent-evaluation, computer-use, memory, rag, reasoning, tool-use
- arXiv categories: cs.AI
- collection score: 19
+- problem: 终局 reward 无法定位 memory 更新中的遗漏、破坏和幻觉，而错误会成为持续系统状态。
+- method: Memory Transition Verifier 为局部更新打分，并用 Transition-Ranked GRPO 优化更新策略。
+- evidence: 相比各错误类型最强 baseline，遗漏、破坏、幻觉分别减少 40.1%、79.1%、50.0%。
+- boundary: verifier 本身仍由冻结 LLM 实现；内容可靠不等于来源可信或具备行动权限。

-## Review Checklist
+## Used In

- Does this paper directly inform Agent architecture, evaluation, memory, tools, safety, coding agents, GUI/browser agents, or multi-agent workflows?
- Does it include a benchmark, dataset, code, or reproducible experimental setup?
- Should it be promoted from `queued` to `skimmed` or `summarized`?
+- [Agent Memory evidence matrix](../../learning/agent-memory/evidence-matrix.md)

 ## Links