Add Agent Memory learning pilot
This commit is contained in:
wuyang
@@ -15,6 +15,7 @@
|
|||||||
- [Glossary](docs/99-glossary.md): 统一术语表
|
- [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
|
- [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
|
- [Papers](papers/README.md): Agent 相关论文、阅读队列和研究洞察
|
||||||
|
- [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
|
- [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
|
- [Projects](projects/README.md): 项目目录、项目记录和复盘
|
||||||
- [Experiments](experiments/README.md): 实验台账、假设、结果和决策
|
- [Experiments](experiments/README.md): 实验台账、假设、结果和决策
|
||||||
|
|||||||
+206
-60
@@ -3517,22 +3517,36 @@
|
|||||||
"collected_at": "2026-07-08",
|
"collected_at": "2026-07-08",
|
||||||
"published_at": "2026-03-12",
|
"published_at": "2026-03-12",
|
||||||
"updated_at": "2026-03-12",
|
"updated_at": "2026-03-12",
|
||||||
"status": "queued",
|
"status": "skimmed",
|
||||||
"relevance": "high",
|
"relevance": "high",
|
||||||
"topics": [
|
"topics": [
|
||||||
"memory",
|
"memory",
|
||||||
"rag"
|
"rag"
|
||||||
],
|
],
|
||||||
"methods": [],
|
"methods": [
|
||||||
"benchmarks": [],
|
"verified-experience-distillation",
|
||||||
|
"instruction-rewriting",
|
||||||
|
"promotion-gate"
|
||||||
|
],
|
||||||
|
"benchmarks": [
|
||||||
|
"CUAD",
|
||||||
|
"HotpotQA"
|
||||||
|
],
|
||||||
"models": [],
|
"models": [],
|
||||||
"datasets": [
|
"datasets": [
|
||||||
"cs.AI"
|
"cs.AI"
|
||||||
],
|
],
|
||||||
"related_concepts": [],
|
"related_concepts": [
|
||||||
|
"procedural-memory",
|
||||||
|
"prompt-compilation"
|
||||||
|
],
|
||||||
"related_jobs": [],
|
"related_jobs": [],
|
||||||
"related_experiments": [],
|
"related_experiments": [
|
||||||
"related_projects": [],
|
"KC-001-agent-memory-pilot"
|
||||||
|
],
|
||||||
|
"related_projects": [
|
||||||
|
"learning/agent-memory"
|
||||||
|
],
|
||||||
"collection_score": "13",
|
"collection_score": "13",
|
||||||
"collection_queries": "language-agent"
|
"collection_queries": "language-agent"
|
||||||
}
|
}
|
||||||
@@ -6913,7 +6927,7 @@
|
|||||||
"collected_at": "2026-07-08",
|
"collected_at": "2026-07-08",
|
||||||
"published_at": "2026-05-08",
|
"published_at": "2026-05-08",
|
||||||
"updated_at": "2026-05-14",
|
"updated_at": "2026-05-14",
|
||||||
"status": "queued",
|
"status": "skimmed",
|
||||||
"relevance": "high",
|
"relevance": "high",
|
||||||
"topics": [
|
"topics": [
|
||||||
"agent-evaluation",
|
"agent-evaluation",
|
||||||
@@ -6924,16 +6938,34 @@
|
|||||||
"reasoning",
|
"reasoning",
|
||||||
"tool-use"
|
"tool-use"
|
||||||
],
|
],
|
||||||
"methods": [],
|
"methods": [
|
||||||
"benchmarks": [],
|
"provenance-dag",
|
||||||
|
"td-lambda",
|
||||||
|
"q-guided-retrieval"
|
||||||
|
],
|
||||||
|
"benchmarks": [
|
||||||
|
"LLAB",
|
||||||
|
"LiveCodeBench",
|
||||||
|
"MMMU-Pro",
|
||||||
|
"ERQA",
|
||||||
|
"GPQA-Diamond",
|
||||||
|
"BFCL"
|
||||||
|
],
|
||||||
"models": [],
|
"models": [],
|
||||||
"datasets": [
|
"datasets": [
|
||||||
"cs.AI"
|
"cs.AI"
|
||||||
],
|
],
|
||||||
"related_concepts": [],
|
"related_concepts": [
|
||||||
|
"memory-credit-assignment",
|
||||||
|
"procedural-memory"
|
||||||
|
],
|
||||||
"related_jobs": [],
|
"related_jobs": [],
|
||||||
"related_experiments": [],
|
"related_experiments": [
|
||||||
"related_projects": [],
|
"KC-001-agent-memory-pilot"
|
||||||
|
],
|
||||||
|
"related_projects": [
|
||||||
|
"learning/agent-memory"
|
||||||
|
],
|
||||||
"collection_score": "20",
|
"collection_score": "20",
|
||||||
"collection_queries": "function-calling"
|
"collection_queries": "function-calling"
|
||||||
}
|
}
|
||||||
@@ -7318,23 +7350,37 @@
|
|||||||
"collected_at": "2026-07-08",
|
"collected_at": "2026-07-08",
|
||||||
"published_at": "2026-05-11",
|
"published_at": "2026-05-11",
|
||||||
"updated_at": "2026-05-11",
|
"updated_at": "2026-05-11",
|
||||||
"status": "queued",
|
"status": "skimmed",
|
||||||
"relevance": "high",
|
"relevance": "high",
|
||||||
"topics": [
|
"topics": [
|
||||||
"agent-evaluation",
|
"agent-evaluation",
|
||||||
"memory",
|
"memory",
|
||||||
"planning"
|
"planning"
|
||||||
],
|
],
|
||||||
"methods": [],
|
"methods": [
|
||||||
"benchmarks": [],
|
"decision-rate-distortion",
|
||||||
|
"k-slot-memory",
|
||||||
|
"certified-memory-split"
|
||||||
|
],
|
||||||
|
"benchmarks": [
|
||||||
|
"LoCoMo",
|
||||||
|
"synthetic-decision-diagnostics"
|
||||||
|
],
|
||||||
"models": [],
|
"models": [],
|
||||||
"datasets": [
|
"datasets": [
|
||||||
"cs.AI"
|
"cs.AI"
|
||||||
],
|
],
|
||||||
"related_concepts": [],
|
"related_concepts": [
|
||||||
|
"decision-sufficient-state",
|
||||||
|
"selective-forgetting"
|
||||||
|
],
|
||||||
"related_jobs": [],
|
"related_jobs": [],
|
||||||
"related_experiments": [],
|
"related_experiments": [
|
||||||
"related_projects": [],
|
"KC-001-agent-memory-pilot"
|
||||||
|
],
|
||||||
|
"related_projects": [
|
||||||
|
"learning/agent-memory"
|
||||||
|
],
|
||||||
"collection_score": "13",
|
"collection_score": "13",
|
||||||
"collection_queries": "language-agent"
|
"collection_queries": "language-agent"
|
||||||
}
|
}
|
||||||
@@ -8094,7 +8140,7 @@
|
|||||||
"collected_at": "2026-07-08",
|
"collected_at": "2026-07-08",
|
||||||
"published_at": "2026-05-14",
|
"published_at": "2026-05-14",
|
||||||
"updated_at": "2026-05-14",
|
"updated_at": "2026-05-14",
|
||||||
"status": "queued",
|
"status": "skimmed",
|
||||||
"relevance": "high",
|
"relevance": "high",
|
||||||
"topics": [
|
"topics": [
|
||||||
"agent-evaluation",
|
"agent-evaluation",
|
||||||
@@ -8102,17 +8148,31 @@
|
|||||||
"memory",
|
"memory",
|
||||||
"tool-use"
|
"tool-use"
|
||||||
],
|
],
|
||||||
"methods": [],
|
"methods": [
|
||||||
"benchmarks": [],
|
"cryptographic-provenance",
|
||||||
|
"derivation-lineage",
|
||||||
|
"sensitive-action-gate"
|
||||||
|
],
|
||||||
|
"benchmarks": [
|
||||||
|
"custom-cross-session-memory-harness",
|
||||||
|
"AgentDojo-bridge"
|
||||||
|
],
|
||||||
"models": [],
|
"models": [],
|
||||||
"datasets": [
|
"datasets": [
|
||||||
"cs.CR",
|
"cs.CR",
|
||||||
"cs.AI"
|
"cs.AI"
|
||||||
],
|
],
|
||||||
"related_concepts": [],
|
"related_concepts": [
|
||||||
|
"memory-provenance",
|
||||||
|
"memory-poisoning"
|
||||||
|
],
|
||||||
"related_jobs": [],
|
"related_jobs": [],
|
||||||
"related_experiments": [],
|
"related_experiments": [
|
||||||
"related_projects": [],
|
"KC-001-agent-memory-pilot"
|
||||||
|
],
|
||||||
|
"related_projects": [
|
||||||
|
"learning/agent-memory"
|
||||||
|
],
|
||||||
"collection_score": "17",
|
"collection_score": "17",
|
||||||
"collection_queries": "agent-memory"
|
"collection_queries": "agent-memory"
|
||||||
}
|
}
|
||||||
@@ -14177,7 +14237,7 @@
|
|||||||
"collected_at": "2026-07-08",
|
"collected_at": "2026-07-08",
|
||||||
"published_at": "2026-06-03",
|
"published_at": "2026-06-03",
|
||||||
"updated_at": "2026-06-03",
|
"updated_at": "2026-06-03",
|
||||||
"status": "queued",
|
"status": "skimmed",
|
||||||
"relevance": "high",
|
"relevance": "high",
|
||||||
"topics": [
|
"topics": [
|
||||||
"agent-evaluation",
|
"agent-evaluation",
|
||||||
@@ -14186,16 +14246,32 @@
|
|||||||
"rag",
|
"rag",
|
||||||
"tool-use"
|
"tool-use"
|
||||||
],
|
],
|
||||||
"methods": [],
|
"methods": [
|
||||||
"benchmarks": [],
|
"agentic-memory-harness",
|
||||||
|
"schema-diagnostics",
|
||||||
|
"active-retrieval"
|
||||||
|
],
|
||||||
|
"benchmarks": [
|
||||||
|
"LoCoMo",
|
||||||
|
"HotpotQA",
|
||||||
|
"AMABench",
|
||||||
|
"ALFWorld"
|
||||||
|
],
|
||||||
"models": [],
|
"models": [],
|
||||||
"datasets": [
|
"datasets": [
|
||||||
"cs.AI"
|
"cs.AI"
|
||||||
],
|
],
|
||||||
"related_concepts": [],
|
"related_concepts": [
|
||||||
|
"schema-commitment",
|
||||||
|
"agentic-retrieval"
|
||||||
|
],
|
||||||
"related_jobs": [],
|
"related_jobs": [],
|
||||||
"related_experiments": [],
|
"related_experiments": [
|
||||||
"related_projects": [],
|
"KC-001-agent-memory-pilot"
|
||||||
|
],
|
||||||
|
"related_projects": [
|
||||||
|
"learning/agent-memory"
|
||||||
|
],
|
||||||
"collection_score": "19",
|
"collection_score": "19",
|
||||||
"collection_queries": "agent-memory"
|
"collection_queries": "agent-memory"
|
||||||
}
|
}
|
||||||
@@ -14980,7 +15056,7 @@
|
|||||||
"collected_at": "2026-07-08",
|
"collected_at": "2026-07-08",
|
||||||
"published_at": "2026-06-04",
|
"published_at": "2026-06-04",
|
||||||
"updated_at": "2026-06-04",
|
"updated_at": "2026-06-04",
|
||||||
"status": "queued",
|
"status": "skimmed",
|
||||||
"relevance": "high",
|
"relevance": "high",
|
||||||
"topics": [
|
"topics": [
|
||||||
"computer-use",
|
"computer-use",
|
||||||
@@ -14989,16 +15065,29 @@
|
|||||||
"rag",
|
"rag",
|
||||||
"tool-use"
|
"tool-use"
|
||||||
],
|
],
|
||||||
"methods": [],
|
"methods": [
|
||||||
"benchmarks": [],
|
"hierarchical-execution-state-tree",
|
||||||
|
"grow-compress-maintain-revise",
|
||||||
|
"error-branch-isolation"
|
||||||
|
],
|
||||||
|
"benchmarks": [
|
||||||
|
"MemoryArena"
|
||||||
|
],
|
||||||
"models": [],
|
"models": [],
|
||||||
"datasets": [
|
"datasets": [
|
||||||
"cs.AI"
|
"cs.AI"
|
||||||
],
|
],
|
||||||
"related_concepts": [],
|
"related_concepts": [
|
||||||
|
"execution-state",
|
||||||
|
"long-horizon-agent"
|
||||||
|
],
|
||||||
"related_jobs": [],
|
"related_jobs": [],
|
||||||
"related_experiments": [],
|
"related_experiments": [
|
||||||
"related_projects": [],
|
"KC-001-agent-memory-pilot"
|
||||||
|
],
|
||||||
|
"related_projects": [
|
||||||
|
"learning/agent-memory"
|
||||||
|
],
|
||||||
"collection_score": "16",
|
"collection_score": "16",
|
||||||
"collection_queries": "agent-memory, rag-agent"
|
"collection_queries": "agent-memory, rag-agent"
|
||||||
}
|
}
|
||||||
@@ -16894,7 +16983,7 @@
|
|||||||
"collected_at": "2026-07-08",
|
"collected_at": "2026-07-08",
|
||||||
"published_at": "2026-06-09",
|
"published_at": "2026-06-09",
|
||||||
"updated_at": "2026-06-09",
|
"updated_at": "2026-06-09",
|
||||||
"status": "queued",
|
"status": "skimmed",
|
||||||
"relevance": "high",
|
"relevance": "high",
|
||||||
"topics": [
|
"topics": [
|
||||||
"agent-evaluation",
|
"agent-evaluation",
|
||||||
@@ -16902,17 +16991,31 @@
|
|||||||
"rag",
|
"rag",
|
||||||
"tool-use"
|
"tool-use"
|
||||||
],
|
],
|
||||||
"methods": [],
|
"methods": [
|
||||||
"benchmarks": [],
|
"topic-documents",
|
||||||
|
"buffered-consolidation",
|
||||||
|
"agentic-retrieval"
|
||||||
|
],
|
||||||
|
"benchmarks": [
|
||||||
|
"MemoryAgentBench",
|
||||||
|
"LongMemEval"
|
||||||
|
],
|
||||||
"models": [],
|
"models": [],
|
||||||
"datasets": [
|
"datasets": [
|
||||||
"cs.AI",
|
"cs.AI",
|
||||||
"cs.CL"
|
"cs.CL"
|
||||||
],
|
],
|
||||||
"related_concepts": [],
|
"related_concepts": [
|
||||||
|
"semantic-memory",
|
||||||
|
"selective-forgetting"
|
||||||
|
],
|
||||||
"related_jobs": [],
|
"related_jobs": [],
|
||||||
"related_experiments": [],
|
"related_experiments": [
|
||||||
"related_projects": [],
|
"KC-001-agent-memory-pilot"
|
||||||
|
],
|
||||||
|
"related_projects": [
|
||||||
|
"learning/agent-memory"
|
||||||
|
],
|
||||||
"collection_score": "18",
|
"collection_score": "18",
|
||||||
"collection_queries": "agent-memory"
|
"collection_queries": "agent-memory"
|
||||||
}
|
}
|
||||||
@@ -17429,7 +17532,7 @@
|
|||||||
"collected_at": "2026-07-08",
|
"collected_at": "2026-07-08",
|
||||||
"published_at": "2026-06-10",
|
"published_at": "2026-06-10",
|
||||||
"updated_at": "2026-06-10",
|
"updated_at": "2026-06-10",
|
||||||
"status": "queued",
|
"status": "skimmed",
|
||||||
"relevance": "high",
|
"relevance": "high",
|
||||||
"topics": [
|
"topics": [
|
||||||
"agent-evaluation",
|
"agent-evaluation",
|
||||||
@@ -17440,18 +17543,33 @@
|
|||||||
"rag",
|
"rag",
|
||||||
"reasoning"
|
"reasoning"
|
||||||
],
|
],
|
||||||
"methods": [],
|
"methods": [
|
||||||
"benchmarks": [],
|
"hierarchical-memory-workspace",
|
||||||
|
"memory-skill-evolution",
|
||||||
|
"rl-retrieval-agent"
|
||||||
|
],
|
||||||
|
"benchmarks": [
|
||||||
|
"ALFWorld",
|
||||||
|
"LoCoMo",
|
||||||
|
"LongMemEval"
|
||||||
|
],
|
||||||
"models": [],
|
"models": [],
|
||||||
"datasets": [
|
"datasets": [
|
||||||
"cs.AI",
|
"cs.AI",
|
||||||
"cs.CL",
|
"cs.CL",
|
||||||
"cs.LG"
|
"cs.LG"
|
||||||
],
|
],
|
||||||
"related_concepts": [],
|
"related_concepts": [
|
||||||
|
"memory-organization",
|
||||||
|
"agentic-retrieval"
|
||||||
|
],
|
||||||
"related_jobs": [],
|
"related_jobs": [],
|
||||||
"related_experiments": [],
|
"related_experiments": [
|
||||||
"related_projects": [],
|
"KC-001-agent-memory-pilot"
|
||||||
|
],
|
||||||
|
"related_projects": [
|
||||||
|
"learning/agent-memory"
|
||||||
|
],
|
||||||
"collection_score": "18",
|
"collection_score": "18",
|
||||||
"collection_queries": "agent-memory"
|
"collection_queries": "agent-memory"
|
||||||
}
|
}
|
||||||
@@ -18977,7 +19095,7 @@
|
|||||||
"collected_at": "2026-07-08",
|
"collected_at": "2026-07-08",
|
||||||
"published_at": "2026-06-12",
|
"published_at": "2026-06-12",
|
||||||
"updated_at": "2026-06-12",
|
"updated_at": "2026-06-12",
|
||||||
"status": "queued",
|
"status": "skimmed",
|
||||||
"relevance": "high",
|
"relevance": "high",
|
||||||
"topics": [
|
"topics": [
|
||||||
"agent-evaluation",
|
"agent-evaluation",
|
||||||
@@ -18986,16 +19104,29 @@
|
|||||||
"reasoning",
|
"reasoning",
|
||||||
"workflow-agent"
|
"workflow-agent"
|
||||||
],
|
],
|
||||||
"methods": [],
|
"methods": [
|
||||||
"benchmarks": [],
|
"budget-matched-baseline",
|
||||||
|
"online-augmentation-audit"
|
||||||
|
],
|
||||||
|
"benchmarks": [
|
||||||
|
"WebArena",
|
||||||
|
"WorkArena-L1"
|
||||||
|
],
|
||||||
"models": [],
|
"models": [],
|
||||||
"datasets": [
|
"datasets": [
|
||||||
"cs.CL"
|
"cs.CL"
|
||||||
],
|
],
|
||||||
"related_concepts": [],
|
"related_concepts": [
|
||||||
|
"memory-cost-accounting",
|
||||||
|
"evaluation-hygiene"
|
||||||
|
],
|
||||||
"related_jobs": [],
|
"related_jobs": [],
|
||||||
"related_experiments": [],
|
"related_experiments": [
|
||||||
"related_projects": [],
|
"KC-001-agent-memory-pilot"
|
||||||
|
],
|
||||||
|
"related_projects": [
|
||||||
|
"learning/agent-memory"
|
||||||
|
],
|
||||||
"collection_score": "14",
|
"collection_score": "14",
|
||||||
"collection_queries": "web-gui-agent"
|
"collection_queries": "web-gui-agent"
|
||||||
}
|
}
|
||||||
@@ -25757,7 +25888,7 @@
|
|||||||
"collected_at": "2026-07-08",
|
"collected_at": "2026-07-08",
|
||||||
"published_at": "2026-06-23",
|
"published_at": "2026-06-23",
|
||||||
"updated_at": "2026-06-23",
|
"updated_at": "2026-06-23",
|
||||||
"status": "queued",
|
"status": "skimmed",
|
||||||
"relevance": "high",
|
"relevance": "high",
|
||||||
"topics": [
|
"topics": [
|
||||||
"agent-evaluation",
|
"agent-evaluation",
|
||||||
@@ -25767,16 +25898,31 @@
|
|||||||
"reasoning",
|
"reasoning",
|
||||||
"tool-use"
|
"tool-use"
|
||||||
],
|
],
|
||||||
"methods": [],
|
"methods": [
|
||||||
"benchmarks": [],
|
"memory-transition-verification",
|
||||||
|
"transition-ranked-grpo",
|
||||||
|
"write-revise-prune"
|
||||||
|
],
|
||||||
|
"benchmarks": [
|
||||||
|
"MemoryAgentBench",
|
||||||
|
"HaluMem",
|
||||||
|
"Mem-alpha"
|
||||||
|
],
|
||||||
"models": [],
|
"models": [],
|
||||||
"datasets": [
|
"datasets": [
|
||||||
"cs.AI"
|
"cs.AI"
|
||||||
],
|
],
|
||||||
"related_concepts": [],
|
"related_concepts": [
|
||||||
|
"memory-consolidation",
|
||||||
|
"transition-verification"
|
||||||
|
],
|
||||||
"related_jobs": [],
|
"related_jobs": [],
|
||||||
"related_experiments": [],
|
"related_experiments": [
|
||||||
"related_projects": [],
|
"KC-001-agent-memory-pilot"
|
||||||
|
],
|
||||||
|
"related_projects": [
|
||||||
|
"learning/agent-memory"
|
||||||
|
],
|
||||||
"collection_score": "19",
|
"collection_score": "19",
|
||||||
"collection_queries": "agent-memory"
|
"collection_queries": "agent-memory"
|
||||||
}
|
}
|
||||||
|
|||||||
+2
-2
@@ -11,8 +11,8 @@
|
|||||||
"unknown": 2
|
"unknown": 2
|
||||||
},
|
},
|
||||||
"papers": {
|
"papers": {
|
||||||
"queued": 969,
|
"queued": 959,
|
||||||
"skimmed": 5
|
"skimmed": 15
|
||||||
},
|
},
|
||||||
"industry": {
|
"industry": {
|
||||||
"analyzed": 8,
|
"analyzed": 8,
|
||||||
|
|||||||
@@ -6,7 +6,7 @@
|
|||||||
|
|
||||||
| ID | Hypothesis | Status | Result | Link |
|
| ID | Hypothesis | Status | Result | Link |
|
||||||
| --- | --- | --- | --- | --- |
|
| --- | --- | --- | --- | --- |
|
||||||
| _TBD_ | _待补充_ | seed | _待补充_ | _待补充_ |
|
| KC-001 | 主题卷宗比论文列表更适合形成可检验认知 | running | 已完成 Agent Memory 第一轮编译，待真实学习反馈 | [记录](knowledge-compilation/2026-07-10-agent-memory-pilot.md) |
|
||||||
|
|
||||||
## Experiment Rules
|
## Experiment Rules
|
||||||
|
|
||||||
@@ -17,4 +17,3 @@
|
|||||||
- 实验结论要写清楚适用边界。
|
- 实验结论要写清楚适用边界。
|
||||||
|
|
||||||
新实验建议复制 [experiment-note](../templates/experiment-note.md)。
|
新实验建议复制 [experiment-note](../templates/experiment-note.md)。
|
||||||
|
|
||||||
|
|||||||
@@ -0,0 +1,154 @@
|
|||||||
|
# KC-001：Agent Memory 主题卷宗试验
|
||||||
|
|
||||||
|
status: running
|
||||||
|
date: 2026-07-10
|
||||||
|
owner: Codex + user
|
||||||
|
|
||||||
|
## 目标
|
||||||
|
|
||||||
|
验证一个核心假设：相较于论文列表、摘要和图谱，把论文编译成“认知地图 + 压缩讲义 + 证据矩阵 + 主动回忆”是否更能帮助人建立可迁移、可检验的领域认知。
|
||||||
|
|
||||||
|
这次不是评估某个 memory 系统，而是评估一种知识加工流程。
|
||||||
|
|
||||||
|
## 为什么选 Agent Memory
|
||||||
|
|
||||||
|
候选主题包括 evaluation、tool use 和 memory。最终选择 memory 的依据：
|
||||||
|
|
||||||
|
- 现有语料中 `agent-memory` primary query 有 103 篇，自动 `memory` 标签约 296 篇，材料足够形成多种方法对照。
|
||||||
|
- 已有 2 篇人工 skimmed 锚点：memory survey 和 EvoMemBench，可以作为问题框架和反例入口。
|
||||||
|
- 主题有清晰演化链：context -> retrieval -> organization -> execution state -> procedural memory -> learned control -> governance。
|
||||||
|
- 它与当前项目“把论文变成长期可调用认知”的目标同构，学习结果可以直接反哺系统设计。
|
||||||
|
|
||||||
|
## 输入
|
||||||
|
|
||||||
|
- 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。
|
||||||
|
- 本轮锚点：12 篇原始论文。
|
||||||
|
|
||||||
|
## 锚点选择规则
|
||||||
|
|
||||||
|
每篇论文必须承担不同的认知角色，避免堆积同类 SOTA：
|
||||||
|
|
||||||
|
1. 一篇统一 taxonomy。
|
||||||
|
2. 一篇统一 benchmark 或强反例。
|
||||||
|
3. 一篇决策理论。
|
||||||
|
4. 至少三种不同表示：主题文档、执行状态、程序记忆。
|
||||||
|
5. 至少一种 learned control / credit assignment。
|
||||||
|
6. 一篇成本与 baseline 反例。
|
||||||
|
7. 写入可靠性和来源治理各一个代表。
|
||||||
|
|
||||||
|
本轮 12 篇：
|
||||||
|
|
||||||
|
| Role | arXiv | Short name |
|
||||||
|
| --- | --- | --- |
|
||||||
|
| taxonomy | 2603.07670 | Memory for Autonomous LLM Agents |
|
||||||
|
| benchmark | 2605.18421 | EvoMemBench |
|
||||||
|
| theory | 2605.10870 | DeMem |
|
||||||
|
| procedural | 2603.15666 | Compiled Memory |
|
||||||
|
| cross-scenario diagnosis | 2606.04315 | AutoMEM |
|
||||||
|
| execution state | 2606.06090 | MAGE |
|
||||||
|
| topic documents | 2606.10677 | Infini Memory |
|
||||||
|
| organize/retrieve split | 2606.11680 | HORMA |
|
||||||
|
| transition reliability | 2606.25161 | TrustMem |
|
||||||
|
| credit assignment | 2605.08374 | MemQ |
|
||||||
|
| budget control | 2606.15017 | Online Skill/Memory Budget Study |
|
||||||
|
| provenance governance | 2605.14421 | MemLineage |
|
||||||
|
|
||||||
|
## 实际流程
|
||||||
|
|
||||||
|
1. 读取本地 corpus summary、reading queue、已有人工笔记和 metadata。
|
||||||
|
2. 比较 memory、evaluation、tool use 的规模与材料成熟度。
|
||||||
|
3. 从约 296 篇 memory 标签论文中选择 12 个互补锚点。
|
||||||
|
4. 通过 arXiv API 校验题目、版本、作者摘要和发布日期。
|
||||||
|
5. 下载 12 篇原始 PDF，用 `pdftotext -layout` 提取可搜索文本。
|
||||||
|
6. 对每篇定位 introduction、method、experiment、ablation、limitation 和 conclusion。
|
||||||
|
7. 先写跨论文问题结构，再把单篇结果放入相应位置，避免让论文目录决定讲义结构。
|
||||||
|
8. 对所有定量主张建立证据矩阵，并同时记录不能外推的范围。
|
||||||
|
9. 把综合认知转换成主动回忆、比较、诊断和设计题。
|
||||||
|
10. 输出 `module.json`，为后续前端学习状态和复习调度保留结构。
|
||||||
|
|
||||||
|
## 模型与工具分工
|
||||||
|
|
||||||
|
本轮没有调用本地 Ollama 做高层综合。原因不是排斥小模型，而是第一版需要先建立质量基线；主题结构、证据冲突和外推边界由 Codex 直接阅读原文后整理。
|
||||||
|
|
||||||
|
后续可交给本地模型的工作：
|
||||||
|
|
||||||
|
- 从新增论文抽取 problem/method/benchmark/limitation 候选字段。
|
||||||
|
- 对锚点论文做章节定位、术语归一和重复检测。
|
||||||
|
- 基于已经确认的讲义生成低风险复习题候选。
|
||||||
|
- 检测新论文是否挑战证据矩阵中的既有判断。
|
||||||
|
|
||||||
|
仍应由强模型或 Codex 负责的工作：
|
||||||
|
|
||||||
|
- 主题边界和锚点选择。
|
||||||
|
- 跨论文因果解释与方法演化。
|
||||||
|
- 证据强弱、实验公平性和外推边界。
|
||||||
|
- 对讲义、测验和系统设计的最终审校。
|
||||||
|
|
||||||
|
## 本轮产物
|
||||||
|
|
||||||
|
- `learning/agent-memory/README.md`
|
||||||
|
- `learning/agent-memory/knowledge-map.md`
|
||||||
|
- `learning/agent-memory/chapter.md`
|
||||||
|
- `learning/agent-memory/evidence-matrix.md`
|
||||||
|
- `learning/agent-memory/active-recall.md`
|
||||||
|
- `learning/agent-memory/module.json`
|
||||||
|
|
||||||
|
## 已观察到的方法经验
|
||||||
|
|
||||||
|
### 1. 自动标签适合召回，不适合直接形成知识结构
|
||||||
|
|
||||||
|
`memory` 标签包含安全、RAG、GUI、推荐等大量邻近论文。它能建立证据池，但不能回答“这个主题最核心的认知冲突是什么”。第一轮必须先选角色互补的锚点。
|
||||||
|
|
||||||
|
### 2. 摘要足以筛选，不足以写可信讲义
|
||||||
|
|
||||||
|
摘要能说明作者想解决什么，却经常省略最重要的限定：baseline 是否公平、prompt 是否更长、结果是否来自自定义 harness、失败集中在哪个子任务。至少需要阅读实验、消融和 limitation。
|
||||||
|
|
||||||
|
### 3. 先搭问题骨架，再放论文
|
||||||
|
|
||||||
|
如果按论文逐篇总结，结果仍然是 12 篇摘要。先确定“状态、表示、控制、学习、治理、评估”这些问题，再让论文作为证据进入，才能形成可迁移认知。
|
||||||
|
|
||||||
|
### 4. 反例和负结果是讲义主干
|
||||||
|
|
||||||
|
EvoMemBench、AutoMEM 诊断和预算研究对认知的贡献不亚于新方法。它们帮助建立三条防误用规则：长上下文必须比较；schema 会造成不可恢复的信息损失；memory 开销必须与 actor budget 公平比较。
|
||||||
|
|
||||||
|
### 5. 原图不是第一版的必要条件
|
||||||
|
|
||||||
|
作者架构图适合解释单篇系统，但第一轮更需要跨论文统一地图。当前使用重新绘制的 Mermaid 关系图，避免让 12 套符号系统增加认知负担。后续只有在公式推导或结构细节无法重画时，再引用原图局部。
|
||||||
|
|
||||||
|
### 6. 测验必须从复述走向设计
|
||||||
|
|
||||||
|
只问“某论文提出了什么”测试的是标题记忆。当前题目按概念、比较诊断、系统设计三级组织；是否能为 coding agent 设计 memory，比记住 HORMA 或 MemQ 的名字更能说明掌握程度。
|
||||||
|
|
||||||
|
## 目前的失败与不足
|
||||||
|
|
||||||
|
- 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。
|
||||||
|
- 还没有把个人答案、卡点和复习历史写入系统。
|
||||||
|
- 还没有让中控 GPT-5.5 对讲义做独立审稿，当前只有一次研究编辑视角。
|
||||||
|
- 尚未把新模块接入现有网页。
|
||||||
|
|
||||||
|
## 下一轮验证
|
||||||
|
|
||||||
|
让用户按 45 分钟入口真实学习一次，并记录：
|
||||||
|
|
||||||
|
- 实际花费时间。
|
||||||
|
- 12 题第一次得分和卡点类型。
|
||||||
|
- 哪些段落仍像论文综述，哪些确实帮助做出判断。
|
||||||
|
- 能否闭卷解释“什么时候不应该使用 memory”。
|
||||||
|
- 能否为自己的 Agent 画出 memory responsibility boundary。
|
||||||
|
|
||||||
|
根据反馈决定第二轮重点：缩短讲义、补图、增加例子，或进入交互式学习界面。
|
||||||
|
|
||||||
|
## 判定标准
|
||||||
|
|
||||||
|
本试验只有满足以下条件才算 `validated`：
|
||||||
|
|
||||||
|
- 用户能在不看材料时说出至少 4 个可迁移设计原则。
|
||||||
|
- 主动回忆前 10 题达到 20/30 以上。
|
||||||
|
- 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 +1,19 @@
|
|||||||
|
# Learning Modules
|
||||||
|
|
||||||
|
这里保存已经从论文集合中“编译”出来的学习模块。论文是证据来源，模块的目标是帮助人建立认知、主动回忆、比较方法并完成设计判断。
|
||||||
|
|
||||||
|
每个模块至少包含：
|
||||||
|
|
||||||
|
- 一个清晰问题，而不是宽泛主题。
|
||||||
|
- 一张认知地图，说明概念与方法之间的关系。
|
||||||
|
- 一篇压缩讲义，区分事实、证据和综合判断。
|
||||||
|
- 一份证据矩阵，保留结论的出处与适用边界。
|
||||||
|
- 一组主动回忆和设计题，用于验证是否真正掌握。
|
||||||
|
- 一份机器可读的 `module.json`，供后续前端和学习进度系统使用。
|
||||||
|
|
||||||
|
## Modules
|
||||||
|
|
||||||
|
| Module | Question | Status | Entry |
|
||||||
|
| --- | --- | --- | --- |
|
||||||
|
| Agent Memory | Agent 怎样把历史压缩成能改善未来决策的状态？ | pilot | [开始学习](agent-memory/README.md) |
|
||||||
|
|
||||||
@@ -0,0 +1,65 @@
|
|||||||
|
# Agent Memory：从存储历史到编译经验
|
||||||
|
|
||||||
|
status: pilot
|
||||||
|
|
||||||
|
## 这次真正要学会什么
|
||||||
|
|
||||||
|
不是记住 12 个系统名字，而是获得一个可迁移判断：
|
||||||
|
|
||||||
|
> Agent memory 不是“多放一些历史记录”，而是在资源、可靠性和安全约束下，把过去压缩为足以改善未来决策的状态。
|
||||||
|
|
||||||
|
完成本模块后，应该能够：
|
||||||
|
|
||||||
|
1. 解释为什么长上下文、向量检索和摘要都不是通用 memory 方案。
|
||||||
|
2. 根据任务选择事实记忆、过程记忆、执行状态或编译指令。
|
||||||
|
3. 设计 write、manage、read、learn、govern 五个环节。
|
||||||
|
4. 用正确 baseline 判断 memory 是否真的带来收益。
|
||||||
|
5. 分析一个 memory 系统可能发生的遗漏、污染、过期和成本问题。
|
||||||
|
|
||||||
|
## 第一次学习：45 分钟
|
||||||
|
|
||||||
|
### 0-5 分钟：先回答
|
||||||
|
|
||||||
|
不要看后文，先写下你的直觉答案：
|
||||||
|
|
||||||
|
- 如果上下文窗口无限大，Agent 还需要 memory 吗？
|
||||||
|
- 一条历史记录什么时候值得被记住？
|
||||||
|
- 相似度最高的记录为什么可能不是最有用的记录？
|
||||||
|
|
||||||
|
### 5-15 分钟：建立地图
|
||||||
|
|
||||||
|
阅读 [认知地图](knowledge-map.md)，重点理解两条轴：
|
||||||
|
|
||||||
|
- 时间：当前任务内还是跨任务。
|
||||||
|
- 内容：知识事实还是执行经验。
|
||||||
|
|
||||||
|
### 15-30 分钟：理解主线
|
||||||
|
|
||||||
|
阅读 [压缩讲义](chapter.md) 的前六节。第一次不要追系统细节，只追踪问题怎样演化：
|
||||||
|
|
||||||
|
`保存历史 -> 检索历史 -> 组织历史 -> 维护执行状态 -> 编译经验 -> 学习记忆策略`
|
||||||
|
|
||||||
|
### 30-38 分钟：检查证据
|
||||||
|
|
||||||
|
阅读 [证据矩阵](evidence-matrix.md)。至少选择一个正面结果和一个反例，确认自己不会把“某篇论文有效”误写成“memory 总是有效”。
|
||||||
|
|
||||||
|
### 38-45 分钟：闭卷输出
|
||||||
|
|
||||||
|
完成 [主动回忆](active-recall.md) 中的前 6 题。答不出来时不要立刻翻答案，先标记卡点属于：概念不清、证据没记住，还是无法迁移到设计。
|
||||||
|
|
||||||
|
## 这个模块的组成
|
||||||
|
|
||||||
|
- [认知地图](knowledge-map.md)：问题空间、方法位置和演化脉络。
|
||||||
|
- [压缩讲义](chapter.md)：从原始论文压缩出的统一解释。
|
||||||
|
- [证据矩阵](evidence-matrix.md)：12 篇锚点论文的主张、证据与边界。
|
||||||
|
- [主动回忆](active-recall.md)：记忆题、比较题、诊断题和设计题。
|
||||||
|
- [module.json](module.json)：供后续学习界面使用的结构化入口。
|
||||||
|
- [试验记录](../../experiments/knowledge-compilation/2026-07-10-agent-memory-pilot.md)：本轮是怎样做出来的，以及下一轮该改什么。
|
||||||
|
|
||||||
|
## 当前边界
|
||||||
|
|
||||||
|
- 当前是第一轮认知编译，不是 Agent Memory 全量综述。
|
||||||
|
- 锚点论文来自现有一年语料中的 12 篇，另外约 300 篇 memory 标签论文暂作证据池。
|
||||||
|
- 已阅读锚点论文原文的关键方法、实验和限制部分，但尚未逐项复现实验。
|
||||||
|
- 论文集中在 2026 年，结论可能受新 benchmark、强模型和评测协议变化影响。
|
||||||
|
|
||||||
@@ -0,0 +1,150 @@
|
|||||||
|
# Agent Memory 主动回忆与设计题
|
||||||
|
|
||||||
|
## 使用规则
|
||||||
|
|
||||||
|
先闭卷回答，再展开参考答案。每题按下面标准自评：
|
||||||
|
|
||||||
|
- `0`：没有答案，或只复述系统名称。
|
||||||
|
- `1`：能说出概念，但不能解释原因或边界。
|
||||||
|
- `2`：能给出机制、证据或反例。
|
||||||
|
- `3`：能迁移到新任务并做出设计判断。
|
||||||
|
|
||||||
|
第一轮不追求满分。真正值得记录的是“我在哪种问题上无法自己推导”。
|
||||||
|
|
||||||
|
## 第一组：核心概念
|
||||||
|
|
||||||
|
### 1. Agent memory 和普通 RAG 的根本区别是什么？
|
||||||
|
|
||||||
|
<details>
|
||||||
|
<summary>参考答案</summary>
|
||||||
|
|
||||||
|
普通 RAG 主要解决外部知识的查询与注入；Agent memory 还要处理由交互产生的状态如何写入、更新、遗忘并影响未来动作。它与 policy 形成反馈环，一次错误写入会改变后续行为，因此还需要 transition verification、credit assignment 和 governance。
|
||||||
|
|
||||||
|
</details>
|
||||||
|
|
||||||
|
### 2. 为什么语义相似的两段历史不一定能被压成同一个 memory state？
|
||||||
|
|
||||||
|
<details>
|
||||||
|
<summary>参考答案</summary>
|
||||||
|
|
||||||
|
因为它们可能要求不同动作。安全合并的判断应是：对当前问题，是否存在一个共同动作，让两段历史的决策损失都在允许范围内。描述相似只是检索信号，不是决策兼容证明。
|
||||||
|
|
||||||
|
</details>
|
||||||
|
|
||||||
|
### 3. 长上下文已经能放下全部历史时，为什么仍可能需要 memory？
|
||||||
|
|
||||||
|
<details>
|
||||||
|
<summary>参考答案</summary>
|
||||||
|
|
||||||
|
可能仍需要降低延迟与 token、减少干扰、跨 session 持久化、支持删除与来源治理、维护执行分支，以及把经验编译成规则。但如果任务短、成本可接受且无需持续学习，长上下文可能就是更好的方案，所以必须作为 baseline。
|
||||||
|
|
||||||
|
</details>
|
||||||
|
|
||||||
|
### 4. 事实记忆和执行状态有什么区别？
|
||||||
|
|
||||||
|
<details>
|
||||||
|
<summary>参考答案</summary>
|
||||||
|
|
||||||
|
事实记忆回答“世界或用户是什么样”，适合主题文档、时间线和结构查询；执行状态回答“任务走到了哪里、哪些约束和分支仍有效”，需要保留动作顺序、子目标边界和错误路径。只用事实检索重建长任务状态容易发生碎片化。
|
||||||
|
|
||||||
|
</details>
|
||||||
|
|
||||||
|
### 5. 什么经验适合被编译进 prompt 或 skill？
|
||||||
|
|
||||||
|
<details>
|
||||||
|
<summary>参考答案</summary>
|
||||||
|
|
||||||
|
跨样本稳定、反复出现、可以验证、能明确改变行为边界的经验。例如固定抽取边界、可靠操作流程或失败恢复规则。短期事实、未经验证的反思和高度状态依赖的决策不适合直接固化。
|
||||||
|
|
||||||
|
</details>
|
||||||
|
|
||||||
|
### 6. 为什么只看最终任务成功不能判断 memory 写入是否可靠？
|
||||||
|
|
||||||
|
<details>
|
||||||
|
<summary>参考答案</summary>
|
||||||
|
|
||||||
|
最终成功可能掩盖中间发生的遗漏、破坏或幻觉，这些错误会留在持久状态中，未来才触发；最终失败也无法定位是哪个 update、retrieval 还是回答步骤造成。因此需要 transition-level 检查和局部归因。
|
||||||
|
|
||||||
|
</details>
|
||||||
|
|
||||||
|
## 第二组：比较与诊断
|
||||||
|
|
||||||
|
### 7. 一个向量库中确实保存了正确答案，但 Agent 仍然回答错误。列出两个不同层面的原因。
|
||||||
|
|
||||||
|
<details>
|
||||||
|
<summary>参考答案</summary>
|
||||||
|
|
||||||
|
表示层可能已经丢失问题所需的动作、时序、因果关系；或者存储保留了细节，但 passive retriever 只返回摘要，没有给 Agent 补查原始证据的机会。还可能是证据已正确取回但 actor 推理失败，这应与 retrieval failure 分开。
|
||||||
|
|
||||||
|
</details>
|
||||||
|
|
||||||
|
### 8. 为什么 stronger retriever 不一定能修复 weak memory manager？
|
||||||
|
|
||||||
|
<details>
|
||||||
|
<summary>参考答案</summary>
|
||||||
|
|
||||||
|
如果写入阶段已经错误合并主题、删除时序、生成错误摘要或遗漏关键谓词，后续检索面对的是被破坏的表示上限。检索只能选择现有内容，不能恢复从未保存的区别。
|
||||||
|
|
||||||
|
</details>
|
||||||
|
|
||||||
|
### 9. 什么时候 memory index 的构建成本可能值得？
|
||||||
|
|
||||||
|
<details>
|
||||||
|
<summary>参考答案</summary>
|
||||||
|
|
||||||
|
需要同时满足：未来问题所需关系能被 schema 表达；很多问题确实需要跨记录聚合；同一索引会被足够多次查询，从而摊销构建成本。否则 raw evidence 加主动工具读取可能更灵活。
|
||||||
|
|
||||||
|
</details>
|
||||||
|
|
||||||
|
### 10. 一个 online memory 方法比 vanilla Agent 成功率高 3%，能否说明有效？还缺什么？
|
||||||
|
|
||||||
|
<details>
|
||||||
|
<summary>参考答案</summary>
|
||||||
|
|
||||||
|
不能。需要计算 memory induction、retrieval、verification 和注入增加的全部 token、延迟和调用；与使用相同预算增加 actor steps 的 baseline 比较；做多次运行并报告方差；检查收益是否集中在特定域以及是否能跨任务复用。
|
||||||
|
|
||||||
|
</details>
|
||||||
|
|
||||||
|
## 第三组：设计题
|
||||||
|
|
||||||
|
### 11. 为一个长期运行的 coding agent 设计 memory
|
||||||
|
|
||||||
|
场景：Agent 每天修改同一仓库，任务跨 session；它需要记住架构约束、失败的修复尝试、当前 issue 进度和可复用命令。请给出最小设计，并说明每种内容放在哪里。
|
||||||
|
|
||||||
|
<details>
|
||||||
|
<summary>参考框架</summary>
|
||||||
|
|
||||||
|
- immutable log：原始终端、代码 diff、测试结果，负责审计和重新解释。
|
||||||
|
- topic documents：模块职责、架构约束、项目惯例，带来源和更新时间。
|
||||||
|
- execution state：当前 issue 的计划、完成步骤、阻塞和失败分支。
|
||||||
|
- compiled skills：经过多次验证的构建、测试、迁移和修复流程。
|
||||||
|
- controller：按任务选择当前状态与相关主题，必要时回查原始日志。
|
||||||
|
- verifier：任何架构事实和规则升级前检查证据；代码执行仍以当前仓库和测试为准。
|
||||||
|
- baseline：与 repo 文档 + 当前 session 长上下文比较成功率、token、延迟和错误恢复。
|
||||||
|
|
||||||
|
</details>
|
||||||
|
|
||||||
|
### 12. 设计一次 memory-vs-long-context 实验
|
||||||
|
|
||||||
|
要求：至少包含任务集、对照组、主要指标、失败分类和安全检查。
|
||||||
|
|
||||||
|
<details>
|
||||||
|
<summary>参考框架</summary>
|
||||||
|
|
||||||
|
- 固定任务：跨 session 事实更新、长程执行、重复流程迁移各一组。
|
||||||
|
- 对照：truncation、full/long context、budget-matched vanilla actor、目标 memory 方法。
|
||||||
|
- 指标：task success、总 token、p95 latency、memory size、过期事实率、错误恢复率。
|
||||||
|
- 失败分类：write omission/corruption、representation loss、retrieval miss、reasoning failure、action failure。
|
||||||
|
- 安全：植入一条外部不可信指令，检查它能否跨 session 触发敏感动作；验证删除请求是否真正传播。
|
||||||
|
- 重复运行：至少多 seed，报告均值和方差，不只报告最好一次。
|
||||||
|
|
||||||
|
</details>
|
||||||
|
|
||||||
|
## 掌握度记录
|
||||||
|
|
||||||
|
| 日期 | 总分 / 36 | 最弱环节 | 能否完成设计题 | 下一次复习 |
|
||||||
|
| --- | ---: | --- | --- | --- |
|
||||||
|
| _待填写_ | | | | |
|
||||||
|
|
||||||
|
建议复习节奏：第一次完成后 1 天、3 天、7 天分别闭卷再答。第二次开始只回答上次低于 2 分的题，并增加一个来自自己项目的新案例。
|
||||||
|
|
||||||
@@ -0,0 +1,189 @@
|
|||||||
|
# 压缩讲义：Agent 怎样从历史中学习
|
||||||
|
|
||||||
|
## 1. Memory 不是仓库，而是受限状态
|
||||||
|
|
||||||
|
最朴素的 Agent memory 是保存聊天记录，然后按相似度找回来。这个模型隐含了两个假设：过去可以被完整保存；只要找到相关文本，模型就能做出更好的动作。最近一年的工作反复证明，这两个假设都不够。
|
||||||
|
|
||||||
|
更有用的形式化来自 POMDP：Agent 看不到完整世界状态，只能用历史构造一个内部状态。设历史为 `h`、当前问题为 `q`、memory 编码器为 `g`，实际决策只能依赖：
|
||||||
|
|
||||||
|
```text
|
||||||
|
m = g(h, q)
|
||||||
|
a = policy(m, q)
|
||||||
|
```
|
||||||
|
|
||||||
|
因此 memory 的目标不是重建全部过去，而是让 `m` 成为当前决策的充分状态。这个视角立刻解释了三个现象：
|
||||||
|
|
||||||
|
- 两段语义相似的历史，可能要求完全不同的行动，因此不能合并。
|
||||||
|
- 两段描述完全不同的历史，如果对当前问题导向同一个最优行动，可以安全压缩。
|
||||||
|
- memory 质量最终要用决策损失衡量，不能只用召回率或摘要相似度衡量。
|
||||||
|
|
||||||
|
[DeMem](https://arxiv.org/abs/2605.10870) 把这个直觉写成 decision rate-distortion。对历史 `h` 采取动作 `a` 的损失是：
|
||||||
|
|
||||||
|
```text
|
||||||
|
Delta_q(h, a) = best_value(h, q) - value(h, q, a)
|
||||||
|
```
|
||||||
|
|
||||||
|
如果一组历史存在一个共同动作，使每段历史的损失都不超过容忍值 `epsilon`，它们才可以共享一个 memory state。这给出了一个非常实用的遗忘原则：
|
||||||
|
|
||||||
|
> 可以忘记描述差异，但不能忘记会改变决策的差异。
|
||||||
|
|
||||||
|
## 2. 为什么长上下文不是答案，但必须作为 baseline
|
||||||
|
|
||||||
|
长上下文保留原始信息，避免了过早摘要和 schema 丢失，所以它经常比复杂 memory 系统更强。[EvoMemBench](https://arxiv.org/abs/2605.18421) 在统一协议下比较 15 种方法，得到三个重要结果：
|
||||||
|
|
||||||
|
1. 长上下文 baseline 仍然很有竞争力。
|
||||||
|
2. memory 在当前上下文不足、任务更难时才更稳定地产生价值。
|
||||||
|
3. 没有一种 memory 形态能同时支配所有任务。
|
||||||
|
|
||||||
|
这不意味着只要扩大窗口即可。上下文越长，推理成本和延迟越高，干扰也越严重；而跨 session 的持久状态、可删除性、来源治理和经验学习也不是上下文窗口本身能解决的。
|
||||||
|
|
||||||
|
正确结论是：长上下文是 memory 实验必须击败的控制组，而不是必须淘汰的旧方案。
|
||||||
|
|
||||||
|
## 3. 相似度为什么经常失效
|
||||||
|
|
||||||
|
向量检索回答的是“什么文本看起来像当前问题”，Agent 真正需要的却可能是：
|
||||||
|
|
||||||
|
- 哪个动作导致了当前状态？
|
||||||
|
- 哪条约束仍然有效？
|
||||||
|
- 上次失败发生在哪个分支？
|
||||||
|
- 哪段经验对未来成功有长期贡献？
|
||||||
|
|
||||||
|
这些关系分别是因果、时间、控制流和价值关系，不能被单一语义距离稳定表达。
|
||||||
|
|
||||||
|
[AutoMEM 的诊断研究](https://arxiv.org/abs/2606.04315) 给出了两个具体失效层：
|
||||||
|
|
||||||
|
- 表示失败：索引 schema 根本没有保存 step id、动作或未来问题需要的谓词，之后再强的检索器也找不回来。
|
||||||
|
- 检索失败：原始存储里仍有答案，但一次被动路由没有把细节交给回答模型，也没有第二次补证据的机会。
|
||||||
|
|
||||||
|
它还给出了“什么时候索引值得做”的两个条件：schema 必须能表达问题所需关系；大量问题必须确实需要跨记录聚合。否则，保留原始文本并让 Agent 用 `grep/read` 主动查找，可能更通用。
|
||||||
|
|
||||||
|
所以结构化不是越多越好。结构化是在写入时对未来问题下注，而主动检索是在查询时再决定需要什么结构。
|
||||||
|
|
||||||
|
## 4. 四类 memory，解决四类失败
|
||||||
|
|
||||||
|
### 4.1 事实与主题文档
|
||||||
|
|
||||||
|
适合用户偏好、人物关系、项目事实和跨 session 知识更新。关键不是把每句话切成孤立向量，而是把相关证据维护成可以修订的主题单元。
|
||||||
|
|
||||||
|
[Infini Memory](https://arxiv.org/abs/2606.10677) 使用 topic documents：新观察先进入缓冲区，再周期性巩固到主题文档；读取时由 Agent 迭代搜索、查看局部行和补充证据。在 MemoryAgentBench 上，其 agentic variant 报告 64.7% overall；维护结构固定时，agentic retrieval 相比 summary+BM25 增加 3.3 点，而移除 split/merge 使 LongMemEval 减少 6.7 点。
|
||||||
|
|
||||||
|
这里的启发不是“文件一定胜过数据库”，而是 memory 需要可维护的语义单元、原始证据链接和多步检查能力。
|
||||||
|
|
||||||
|
### 4.2 当前执行状态
|
||||||
|
|
||||||
|
长程工具任务的主要问题常常不是忘记一个事实，而是丢失当前控制状态：已经完成哪些子目标、哪些路径失败、哪些动作改变了后续约束。
|
||||||
|
|
||||||
|
[MAGE](https://arxiv.org/abs/2606.06090) 用两层树表示执行状态：底层保存 action-observation，顶层在子目标边界压缩；当前状态来自 root-to-current path。四个操作形成闭环：
|
||||||
|
|
||||||
|
- `Grow`：追加新的动作与观察。
|
||||||
|
- `Compress`：在边界把完成片段压缩为子目标状态。
|
||||||
|
- `Maintain`：在摘要成为可信状态前检查遗漏和执行错误。
|
||||||
|
- `Revise`：回到错误边界，从新分支继续，隔离错误轨迹。
|
||||||
|
|
||||||
|
这个结构解决的是控制流连续性，而不是文本相似性。论文在 MemoryArena 报告相对 baseline 平均提高 7.8 到 20.4 个百分点，同时比长上下文减少 55.1% token。
|
||||||
|
|
||||||
|
### 4.3 稳定规则与技能
|
||||||
|
|
||||||
|
如果同一种错误反复发生，把旧案例再次塞进上下文可能不如直接改变 Agent 的行为规则。
|
||||||
|
|
||||||
|
[Compiled Memory](https://arxiv.org/abs/2603.15666) 的 Atlas 把经过验证的经验写成 system prompt 中的子规则：memory 是 distillation，读取是 instruction rewriting。其 CUAD 实验报告 token F1 增加 8.7 个百分点，HotpotQA joint F1 增加 3.16 个百分点；同一份从 GPT-4o 错误编译的 prompt 在 Claude Sonnet 4.5 上仍增加 2.31 个百分点。
|
||||||
|
|
||||||
|
但这个结果有重要限制：演化 prompt 从约 960 增至 1570 token，论文没有完成等长度通用内容控制；而训练信号没有覆盖的目标不会自然改善。这说明程序记忆适合稳定、重复、可验证的任务规则，不适合不断变化的事实。
|
||||||
|
|
||||||
|
### 4.4 有价值的经验链
|
||||||
|
|
||||||
|
一次成功可能依赖更早的记忆，而那条记忆又帮助生成了后续记忆。只奖励最后一次被直接检索的记录，会错误分配信用。
|
||||||
|
|
||||||
|
[MemQ](https://arxiv.org/abs/2605.08374) 记录 memory provenance DAG，并把 TD error 沿祖先链反向传播：
|
||||||
|
|
||||||
|
```text
|
||||||
|
delta(m) = reward + gamma * Q(new_memory) - Q(m)
|
||||||
|
credit(ancestor) += alpha * (gamma * lambda)^depth * delta(m)
|
||||||
|
```
|
||||||
|
|
||||||
|
它在多步任务上的提升大于单步任务，说明“结构距离”可以比单纯时间距离更适合经验归因。不过它假设 memory 单调增长，尚未解决巩固和删除，维护 DAG 也会持续增加成本。
|
||||||
|
|
||||||
|
## 5. 写入比读取更危险
|
||||||
|
|
||||||
|
检索错了一次通常影响一个回答；持久 memory 写错一次，可能在未来反复被召回，成为系统状态污染。
|
||||||
|
|
||||||
|
[TrustMem](https://arxiv.org/abs/2606.25161) 把 memory update 视为可审计 transition，动作空间是 `WRITE / REVISE / PRUNE`。verifier 从三个维度检查每次更新：
|
||||||
|
|
||||||
|
- coverage：新输入中的重要信息是否保留。
|
||||||
|
- preservation：原有有效信息是否被无依据删除或扭曲。
|
||||||
|
- faithfulness：新写内容是否有当前输入或旧状态支持。
|
||||||
|
|
||||||
|
这比只看最终问答正确率更合理，因为正确答案也可能掩盖中间产生的危险状态。论文报告，相比各错误类型的最强 baseline，遗漏、破坏和幻觉分别减少 40.1%、79.1% 和 50.0%。
|
||||||
|
|
||||||
|
但“内容真实”仍不等于“有权支持行动”。外部网页中的恶意指令即使被准确保存，也不应在未来授权转账。[MemLineage](https://arxiv.org/abs/2605.14421) 因此把来源签名、派生 DAG 和敏感动作 gate 结合起来：一条行动理由如果沿派生链来自不可信源，就不能直接执行。
|
||||||
|
|
||||||
|
它的确定性机制隔离实验中，三类攻击的 ASR 均降为零，但这个结果来自自定义 cross-session harness；headline 评估没有覆盖完整自适应攻击，且作者明确假设模型权重和推理路径可信。它支持“来源链必须进入执行策略”这个方向，但不能被读成 memory 安全已解决。
|
||||||
|
|
||||||
|
## 6. Memory controller 才是核心能力
|
||||||
|
|
||||||
|
写入什么、怎样组织、什么时候读取、如何补证据和何时遗忘，本质上都是控制策略。当前出现了三种控制方式：
|
||||||
|
|
||||||
|
| 控制方式 | 优点 | 代价 |
|
||||||
|
| --- | --- | --- |
|
||||||
|
| 固定规则 | 可预测、可调试、成本低 | 难适应任务与状态变化 |
|
||||||
|
| LLM 工具调用 | 灵活，可按问题主动查找 | 多轮调用昂贵，受模型能力影响 |
|
||||||
|
| 训练控制器 | 能从结果优化策略 | 训练和 credit assignment 复杂，容易过拟合 benchmark |
|
||||||
|
|
||||||
|
[HORMA](https://arxiv.org/abs/2606.11680) 提供了一个值得保留的分工：高能力模型负责异步 memory organization，轻量模型学习低延迟 retrieval。论文中，Qwen 3.5 4B retriever 仅在 LoCoMo 训练，却能迁移到 ALFWorld 和 LongMemEval；但弱模型如果一开始就把 memory 组织坏了，更强的 retriever 也补不回来。
|
||||||
|
|
||||||
|
这正好说明我们自己的模型分工原则：小模型适合局部、可验证、低风险操作；跨论文压缩、结构设计和关键判断需要强模型或人工研究编辑承担。
|
||||||
|
|
||||||
|
## 7. 怎样证明 memory 值得存在
|
||||||
|
|
||||||
|
Memory 论文最容易犯的错误，是只比较成功率，不计算 memory 模块自己消耗的 token、延迟和额外调用。
|
||||||
|
|
||||||
|
[在线 memory/skill 的预算研究](https://arxiv.org/abs/2606.15017) 把相同预算交给 vanilla actor 增加观察和行动步数。在 WebArena 的三个域、三个模型上，budget-matched vanilla baseline 的平均成功率都高于三个在线 augmentation 方法，并且通常 token 更少；在 WorkArena-L1 上也保持竞争力。
|
||||||
|
|
||||||
|
这项结果不能推广到所有 memory：它研究的是在线模块，离线编译后被大量复用的技能需要另一种摊销方式。但它给出了必须遵守的评测卫生：
|
||||||
|
|
||||||
|
1. 报告所有模块的 token 和延迟，不只报告主 Agent。
|
||||||
|
2. 与长上下文和 budget-matched vanilla actor 比较。
|
||||||
|
3. 多次运行并报告方差。
|
||||||
|
4. 分开评估写入、组织、读取、回答和动作执行的错误。
|
||||||
|
5. 测试更新、冲突、遗忘、污染和跨任务泛化，不只测静态 recall。
|
||||||
|
|
||||||
|
## 8. 综合结论：一个可工作的参考架构
|
||||||
|
|
||||||
|
对需要长期运行、可学习、可审计的 Agent，可以从下面的最小架构开始：
|
||||||
|
|
||||||
|
```mermaid
|
||||||
|
flowchart TB
|
||||||
|
E["Immutable event log"] --> V["Transition verifier"]
|
||||||
|
V --> B["Current buffer"]
|
||||||
|
B --> T["Topic documents"]
|
||||||
|
B --> X["Execution-state tree"]
|
||||||
|
T --> K["Compiled rules / skills"]
|
||||||
|
X --> K
|
||||||
|
Q["Current task"] --> C["Memory controller"]
|
||||||
|
T --> C
|
||||||
|
X --> C
|
||||||
|
K --> C
|
||||||
|
C --> D["Minimal sufficient context"]
|
||||||
|
D --> A["Agent action"]
|
||||||
|
A --> O["Outcome and feedback"]
|
||||||
|
O --> L["Credit and consolidation"]
|
||||||
|
L --> V
|
||||||
|
P["Provenance / time / principal / policy"] --> V
|
||||||
|
P --> C
|
||||||
|
P --> A
|
||||||
|
```
|
||||||
|
|
||||||
|
这不是要求第一天实现全部层。它给出的是责任边界：
|
||||||
|
|
||||||
|
- 原始日志负责可恢复和审计。
|
||||||
|
- topic documents 负责可维护事实。
|
||||||
|
- execution tree 负责长任务状态。
|
||||||
|
- compiled rules 负责稳定行为改变。
|
||||||
|
- controller 负责按任务选择最小充分上下文。
|
||||||
|
- verifier 与 provenance 负责阻止错误和不可信状态永久化。
|
||||||
|
- outcome loop 负责让 memory 从使用结果中学习。
|
||||||
|
|
||||||
|
最终设计原则可以压成一句话：
|
||||||
|
|
||||||
|
> 先判断未来决策需要保留什么区别，再选择存储结构；先用公平 baseline 证明价值，再让 memory 获得长期权限。
|
||||||
|
|
||||||
@@ -0,0 +1,57 @@
|
|||||||
|
# Agent Memory 证据矩阵
|
||||||
|
|
||||||
|
## 怎么使用这张表
|
||||||
|
|
||||||
|
这不是论文排行榜。每篇只承担一个认知角色：定义问题、提出机制、给出反例或补上治理边界。数字按论文原文记录，尚未由本项目独立复现。
|
||||||
|
|
||||||
|
| 认知角色 | 论文 | 核心主张 | 主要证据 | 不能推出什么 |
|
||||||
|
| --- | --- | --- | --- | --- |
|
||||||
|
| 总体框架 | [Memory for Autonomous LLM Agents](https://arxiv.org/abs/2603.07670) | Memory 是 write-manage-read loop；应同时看时间范围、表示基底和控制策略 | 汇总 2022 至 2026 年初机制与 benchmark，提出 utility、efficiency、adaptivity、faithfulness、governance 五类目标 | 综述中的跨论文数字不是统一协议下的直接比较 |
|
||||||
|
| 通用反例 | [EvoMemBench](https://arxiv.org/abs/2605.18421) | Memory 价值依赖任务范围与内容类型，没有通用最优形态 | 在统一协议下比较 15 种 memory 方法和长上下文 baseline；长上下文仍有竞争力，知识任务偏 retrieval，执行任务偏 procedural/long-term memory | 不能据此认为 memory 无价值；它在上下文不足和困难任务中更有帮助 |
|
||||||
|
| 决策理论 | [Remember the Decision, Not the Description](https://arxiv.org/abs/2605.10870) | 应保留影响决策的历史差异，而不是追求描述保真 | 在 LoCoMo 子集上，描述相似度对 evidence compatibility 的 Spearman 相关仅 0.103；相同预算下 DeMem 的 gold evidence recall 为 83%，描述检索为 66% | 理论主要在 contextual bandit 与有限 memory state 设定下建立，不等于所有 Agent 控制问题已被形式化 |
|
||||||
|
| 行为编译 | [Compiled Memory](https://arxiv.org/abs/2603.15666) | 经验可以被验证后编译进指令，使 memory 直接改变行为 | CUAD token F1 +8.7pp；HotpotQA joint F1 +3.16pp；同一 evolved prompt 在 Claude Sonnet 4.5 上 +2.31pp | evolved prompt 更长，缺少等长度通用内容 control；只会改善训练信号明确覆盖的目标 |
|
||||||
|
| 跨场景诊断 | [Exploring Cross-Scenario Generality / AutoMEM](https://arxiv.org/abs/2606.04315) | 结构化索引可能在写入时丢失未来问题需要的谓词；Agent 主动读取更通用 | 比较 8 个系统和一个 harness、覆盖 5 类场景；AutoMEM 在 LoCoMo 为 67.3，对 long context 的 61.5；ALFWorld 中 golden procedure 71.7、最好 memory 43.3、GRPO actor 86.7 | 主动 harness 并非总是最便宜；动态任务的上限可能来自 actor policy，而不是 memory 结构 |
|
||||||
|
| 执行状态 | [MAGE](https://arxiv.org/abs/2606.06090) | 长任务需要 root-to-current execution state 和错误分支隔离，而非相似记录集合 | MemoryArena 上平均 success rate 相对 baseline +7.8 至 +20.4pp；相对长上下文 token -55.1% | 结果集中在 interdependent long-horizon task，不能直接外推到事实问答与个性化记忆 |
|
||||||
|
| 可维护事实 | [Infini Memory](https://arxiv.org/abs/2606.10677) | Topic documents、周期巩固与多步读取可以共同改善长期记忆 | MemoryAgentBench overall 64.7；LongMemEval 上 summary-only 41.7、summary+BM25 76.0、agentic 79.3；去掉 split/merge 后从 76.0 降到 69.3 | 多跳 selective forgetting 仍明显困难；agentic retrieval 增加调用和维护成本 |
|
||||||
|
| 组织与读取分工 | [HORMA](https://arxiv.org/abs/2606.11680) | 高层组织和低层检索应分开优化；文件层级可作为可操作 workspace | ALFWorld strict context 下达到 56.7/73.9 SR；LoCoMo 和 LongMemEval 上优于列出的 baselines；小型 RL retriever 有跨域迁移 | 论文依赖高能力模型进行 memory construction；组织错误无法被更强检索完全修复 |
|
||||||
|
| 写入可靠性 | [TrustMem](https://arxiv.org/abs/2606.25161) | 应在每次 memory transition 检查 coverage、preservation、faithfulness | MemoryAgentBench 相对最强 baseline +6.5；HaluMem extraction +12.14 F1；遗漏、破坏、幻觉分别减少 40.1%、79.1%、50.0% | verifier 是冻结 LLM，仍可能有偏差；提高 transition 可靠性不等于来源可信或有行动权限 |
|
||||||
|
| 经验价值归因 | [MemQ](https://arxiv.org/abs/2605.08374) | Memory 的长期价值应沿 provenance DAG 反向归因 | 六个 benchmark 上领先或并列；相对单步 MemRL，在深链任务最高 +5.7pp，单步任务差距可低至 0.77pp | 假设 memory 单调增长，尚未处理 consolidation/deletion；DAG 和 BFS 带来持续开销 |
|
||||||
|
| 公平成本控制 | [Are Online Skill and Memory Modules Always Worth Their Tokens?](https://arxiv.org/abs/2606.15017) | Memory/skill 的收益必须与把相同预算交给 actor 的 baseline 比较 | WebArena 三域三模型中 Vanilla-IB 平均 SR 均领先三个 augmentation 方法，通常 token 更少；WorkArena-L1 仍在 Pareto frontier | 结论针对 online augmentation；离线编译并重复摊销的 memory 需要不同成本模型 |
|
||||||
|
| 来源与执行权限 | [MemLineage](https://arxiv.org/abs/2605.14421) | 内容签名不够，派生链中的不可信来源也应阻止敏感行动 | 自定义确定性 harness 中，三类攻击只有 lineage 方案全部达到 0 ASR；报告亚毫秒级单操作开销 | harness 为机制隔离而非完整公共 benchmark；headline 未覆盖完整 adaptive attack，并假设模型与推理路径可信 |
|
||||||
|
|
||||||
|
## 跨论文一致结论
|
||||||
|
|
||||||
|
下面这些判断至少得到两类不同证据支持：
|
||||||
|
|
||||||
|
### 相似度不是 memory 的统一目标
|
||||||
|
|
||||||
|
- DeMem 从决策损失说明描述相似不等于行为兼容。
|
||||||
|
- AutoMEM 显示 schema 和 passive retrieval 会丢失动作、时序与因果信息。
|
||||||
|
- MAGE 显示长任务需要保持执行路径，而不是拼接相似片段。
|
||||||
|
|
||||||
|
### Memory construction 不能被当作低价值预处理
|
||||||
|
|
||||||
|
- Infini Memory 的维护消融下降大于 agentic retrieval 的增益。
|
||||||
|
- HORMA 显示弱 manager 组织出的 workspace 无法由强 retriever 修复。
|
||||||
|
- TrustMem 显示一次错误 transition 会成为长期状态污染。
|
||||||
|
|
||||||
|
### Memory 的收益必须带着成本与 baseline 阅读
|
||||||
|
|
||||||
|
- EvoMemBench 显示长上下文 baseline 仍然强。
|
||||||
|
- 在线预算研究显示，额外 actor steps 能吃掉 augmentation 的表面收益。
|
||||||
|
- HORMA、Infini Memory 和 AutoMEM 都显示多步 agentic retrieval 存在质量与调用成本交换。
|
||||||
|
|
||||||
|
### 自我进化受限于反馈质量和 actor 能力
|
||||||
|
|
||||||
|
- Compiled Memory 明确表现出 training-signal constraint。
|
||||||
|
- MemQ 需要 provenance 才能处理多步 credit assignment。
|
||||||
|
- AutoMEM 的 ALFWorld 诊断显示文本 memory 无法替代 state-dependent policy learning。
|
||||||
|
|
||||||
|
## 当前仍不确定
|
||||||
|
|
||||||
|
- 在同一强模型、同一成本预算和同一任务集下，topic documents、execution tree、compiled instruction 的最优组合是什么。
|
||||||
|
- memory update verifier 是否能在真实长周期运行中抵抗 verifier 自身漂移和系统性偏差。
|
||||||
|
- 主动检索的收益何时足以覆盖额外模型调用；是否可由小模型或确定性工具稳定替代大模型步骤。
|
||||||
|
- 来源 lineage、语义正确性和真实权限系统如何以较低复杂度组合。
|
||||||
|
- 当模型能力持续提升时，哪些 memory scaffolding 会失去价值，哪些仍然是系统层必需能力。
|
||||||
|
|
||||||
@@ -0,0 +1,109 @@
|
|||||||
|
# Agent Memory 认知地图
|
||||||
|
|
||||||
|
## 一句话地图
|
||||||
|
|
||||||
|
Agent memory 的核心任务是：在不能把全部历史交给模型的情况下，保留足以支持未来正确决策的信息结构。
|
||||||
|
|
||||||
|
```mermaid
|
||||||
|
flowchart LR
|
||||||
|
O["观察与轨迹"] --> W["写入判断"]
|
||||||
|
W --> R["原始证据层"]
|
||||||
|
W --> S["语义与主题层"]
|
||||||
|
W --> P["程序与技能层"]
|
||||||
|
R --> X["当前执行状态"]
|
||||||
|
S --> X
|
||||||
|
P --> X
|
||||||
|
X --> A["行动与回答"]
|
||||||
|
A --> F["结果与反馈"]
|
||||||
|
F --> C["价值归因与巩固"]
|
||||||
|
C --> W
|
||||||
|
G["来源、时效与权限"] --> W
|
||||||
|
G --> X
|
||||||
|
```
|
||||||
|
|
||||||
|
这张图里，存储只是中间环节。真正困难的是写入、选择、更新、归因和治理。
|
||||||
|
|
||||||
|
## 两条最重要的区分轴
|
||||||
|
|
||||||
|
EvoMemBench 把问题拆成两条轴，这比“短期/长期 memory”更能指导设计。
|
||||||
|
|
||||||
|
| | 知识导向 | 执行导向 |
|
||||||
|
| --- | --- | --- |
|
||||||
|
| 当前任务内 | 当前问题需要的事实、约束、证据 | 当前计划、子目标、工具结果、失败分支 |
|
||||||
|
| 跨任务 | 用户偏好、领域规则、稳定事实 | 可复用流程、技能、失败教训、策略价值 |
|
||||||
|
|
||||||
|
由此得到四个问题：
|
||||||
|
|
||||||
|
1. 当前任务中，我需要保留哪些事实？
|
||||||
|
2. 当前任务中，我需要维持怎样的执行状态？
|
||||||
|
3. 跨任务后，哪些知识仍然有效？
|
||||||
|
4. 哪些经验应该变成可以复用的行为规则？
|
||||||
|
|
||||||
|
## 四种核心产物
|
||||||
|
|
||||||
|
| 产物 | 保存什么 | 最适合的任务 | 主要风险 |
|
||||||
|
| --- | --- | --- | --- |
|
||||||
|
| 原始证据 | 完整对话、工具结果、轨迹 | 审计、细节恢复、重新解释 | 太长、噪声多、检索困难 |
|
||||||
|
| 语义文档 | 事实、主题、关系、时间线 | 多 session 问答、偏好与知识更新 | 过期、冲突、schema 不匹配 |
|
||||||
|
| 执行状态 | 当前路径、子目标、分支和约束 | 长程工具任务、GUI、编码与规划 | 错误状态持续传播 |
|
||||||
|
| 程序知识 | 规则、技能、指令和策略 | 重复任务、自我改进 | 过拟合、错误经验被固化 |
|
||||||
|
|
||||||
|
一个完整系统通常需要它们共存，而不是选一个数据库解决全部问题。
|
||||||
|
|
||||||
|
## 方法演化
|
||||||
|
|
||||||
|
```mermaid
|
||||||
|
flowchart TD
|
||||||
|
A["把历史塞进上下文"] --> B["按相似度检索记录"]
|
||||||
|
B --> C["结构化主题、图和层级"]
|
||||||
|
C --> D["Agent 主动导航与验证证据"]
|
||||||
|
D --> E["把轨迹维护成执行状态"]
|
||||||
|
E --> F["把经验编译为规则和技能"]
|
||||||
|
F --> H["从结果学习写入、检索和遗忘策略"]
|
||||||
|
H --> I["增加 transition 验证、来源和权限治理"]
|
||||||
|
```
|
||||||
|
|
||||||
|
这不是严格时间线，而是问题不断暴露后的能力叠加：
|
||||||
|
|
||||||
|
- 上下文解决可见性，但产生成本和注意力稀释。
|
||||||
|
- 检索解决容量，但相似不等于对决策有用。
|
||||||
|
- 结构解决关系和聚合，但 schema 可能提前丢失未来问题需要的信息。
|
||||||
|
- Agentic retrieval 延迟结构承诺，让模型按问题寻找证据，但增加推理成本。
|
||||||
|
- 执行状态解决长任务中的因果连续性和错误隔离。
|
||||||
|
- 编译记忆把反复出现的经验变成行为改变，而不只是再次展示历史。
|
||||||
|
- 学习控制策略解决“哪些记忆长期有价值”，同时引入 credit assignment 问题。
|
||||||
|
- 治理层防止错误或不可信内容变成永久系统状态。
|
||||||
|
|
||||||
|
## 五个设计环节
|
||||||
|
|
||||||
|
### Write
|
||||||
|
|
||||||
|
决定候选信息是否值得越过持久化边界。需要处理信息增量、重复、置信度、来源和未来用途。
|
||||||
|
|
||||||
|
### Manage
|
||||||
|
|
||||||
|
执行合并、拆分、更新、遗忘、冲突处理和版本维护。这里决定 memory 会不会随时间腐烂。
|
||||||
|
|
||||||
|
### Read
|
||||||
|
|
||||||
|
根据当前问题选择最小但充分的上下文。读取既可以是一次检索，也可以是 Agent 多步搜索、检查和补证据。
|
||||||
|
|
||||||
|
### Learn
|
||||||
|
|
||||||
|
根据任务结果更新哪些记忆有用、哪些结构有效、哪些经验应晋升为规则。关键难题是把结果正确归因到先前记忆。
|
||||||
|
|
||||||
|
### Govern
|
||||||
|
|
||||||
|
记录来源、时间、主体、权限和派生链；敏感行动不能只因为某条记忆“看起来合理”就执行。
|
||||||
|
|
||||||
|
## 一眼判断方法
|
||||||
|
|
||||||
|
面对一个新任务，按顺序问：
|
||||||
|
|
||||||
|
1. 历史是否真的超过有效上下文，或者会跨 session？如果否，先用长上下文 baseline。
|
||||||
|
2. 任务失败来自忘记事实，还是丢失执行状态？两者需要不同结构。
|
||||||
|
3. 是否存在可重复利用的稳定模式？如果有，考虑程序记忆或编译指令。
|
||||||
|
4. 未来查询类型能否提前定义？如果不能，保留原始证据并支持主动导航。
|
||||||
|
5. memory 模块消耗的 token 和延迟，是否比把预算给基础 Agent 更值得？
|
||||||
|
6. 写错、写旧或写入恶意内容后，系统能否追踪、撤销并阻止敏感行动？
|
||||||
|
|
||||||
@@ -0,0 +1,56 @@
|
|||||||
|
{
|
||||||
|
"id": "agent-memory-pilot",
|
||||||
|
"title": "Agent Memory: 从存储历史到编译经验",
|
||||||
|
"status": "pilot",
|
||||||
|
"version": 1,
|
||||||
|
"created_at": "2026-07-10",
|
||||||
|
"estimated_minutes": 45,
|
||||||
|
"core_question": "Agent 怎样在资源、可靠性和安全约束下，把历史压缩成能改善未来决策的状态？",
|
||||||
|
"learning_outcomes": [
|
||||||
|
"区分长上下文、检索记忆、执行状态和程序记忆",
|
||||||
|
"根据任务选择合适的 memory 表示与控制策略",
|
||||||
|
"设计 write、manage、read、learn、govern 闭环",
|
||||||
|
"使用 budget-matched baseline 和分层错误指标评估 memory",
|
||||||
|
"识别持久 memory 的遗漏、污染、过期和权限风险"
|
||||||
|
],
|
||||||
|
"artifacts": [
|
||||||
|
{"type": "map", "path": "knowledge-map.md"},
|
||||||
|
{"type": "chapter", "path": "chapter.md"},
|
||||||
|
{"type": "evidence", "path": "evidence-matrix.md"},
|
||||||
|
{"type": "recall", "path": "active-recall.md"}
|
||||||
|
],
|
||||||
|
"anchor_papers": [
|
||||||
|
"2603.07670",
|
||||||
|
"2605.18421",
|
||||||
|
"2605.10870",
|
||||||
|
"2603.15666",
|
||||||
|
"2606.04315",
|
||||||
|
"2606.06090",
|
||||||
|
"2606.10677",
|
||||||
|
"2606.11680",
|
||||||
|
"2606.25161",
|
||||||
|
"2605.08374",
|
||||||
|
"2606.15017",
|
||||||
|
"2605.14421"
|
||||||
|
],
|
||||||
|
"concepts": [
|
||||||
|
"decision-sufficient-state",
|
||||||
|
"write-manage-read-loop",
|
||||||
|
"episodic-memory",
|
||||||
|
"semantic-memory",
|
||||||
|
"procedural-memory",
|
||||||
|
"execution-state",
|
||||||
|
"agentic-retrieval",
|
||||||
|
"memory-consolidation",
|
||||||
|
"credit-assignment",
|
||||||
|
"provenance",
|
||||||
|
"selective-forgetting",
|
||||||
|
"budget-matched-evaluation"
|
||||||
|
],
|
||||||
|
"assessment": {
|
||||||
|
"path": "active-recall.md",
|
||||||
|
"question_count": 12,
|
||||||
|
"max_score": 36,
|
||||||
|
"review_intervals_days": [1, 3, 7]
|
||||||
|
}
|
||||||
|
}
|
||||||
+18
-16
@@ -12,47 +12,49 @@ source: arxiv
|
|||||||
collected_at: 2026-07-08
|
collected_at: 2026-07-08
|
||||||
published_at: 2026-03-12
|
published_at: 2026-03-12
|
||||||
updated_at: 2026-03-12
|
updated_at: 2026-03-12
|
||||||
status: queued
|
status: skimmed
|
||||||
relevance: high
|
relevance: high
|
||||||
topics:
|
topics:
|
||||||
- memory
|
- memory
|
||||||
- rag
|
- rag
|
||||||
methods:
|
methods:
|
||||||
-
|
- verified-experience-distillation
|
||||||
|
- instruction-rewriting
|
||||||
|
- promotion-gate
|
||||||
benchmarks:
|
benchmarks:
|
||||||
-
|
- CUAD
|
||||||
|
- HotpotQA
|
||||||
models:
|
models:
|
||||||
-
|
-
|
||||||
datasets:
|
datasets:
|
||||||
- cs.AI
|
- cs.AI
|
||||||
related_concepts:
|
related_concepts:
|
||||||
-
|
- procedural-memory
|
||||||
|
- prompt-compilation
|
||||||
related_jobs:
|
related_jobs:
|
||||||
-
|
-
|
||||||
related_experiments:
|
related_experiments:
|
||||||
-
|
- KC-001-agent-memory-pilot
|
||||||
related_projects:
|
related_projects:
|
||||||
-
|
- learning/agent-memory
|
||||||
collection_score: 13
|
collection_score: 13
|
||||||
collection_queries: language-agent
|
collection_queries: language-agent
|
||||||
---
|
---
|
||||||
|
|
||||||
## One-line Takeaway
|
## 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
|
- problem: 有更多历史不代表 Agent 知道哪些经验应改变未来行为。
|
||||||
- inferred topics: memory, rag
|
- method: Atlas 用多层 promotion 和 verification 把经验写成 system prompt 子规则。
|
||||||
- arXiv categories: cs.AI
|
- evidence: CUAD token F1 +8.7pp，HotpotQA joint F1 +3.16pp；跨模型复用同一 prompt 时 +2.31pp。
|
||||||
- collection score: 13
|
- 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?
|
- [Agent Memory evidence matrix](../../learning/agent-memory/evidence-matrix.md)
|
||||||
- Does it include a benchmark, dataset, code, or reproducible experimental setup?
|
|
||||||
- Should it be promoted from `queued` to `skimmed` or `summarized`?
|
|
||||||
|
|
||||||
## Links
|
## Links
|
||||||
|
|
||||||
|
|||||||
+22
-16
@@ -12,7 +12,7 @@ source: arxiv
|
|||||||
collected_at: 2026-07-08
|
collected_at: 2026-07-08
|
||||||
published_at: 2026-05-08
|
published_at: 2026-05-08
|
||||||
updated_at: 2026-05-14
|
updated_at: 2026-05-14
|
||||||
status: queued
|
status: skimmed
|
||||||
relevance: high
|
relevance: high
|
||||||
topics:
|
topics:
|
||||||
- agent-evaluation
|
- agent-evaluation
|
||||||
@@ -23,41 +23,47 @@ topics:
|
|||||||
- reasoning
|
- reasoning
|
||||||
- tool-use
|
- tool-use
|
||||||
methods:
|
methods:
|
||||||
-
|
- provenance-dag
|
||||||
|
- td-lambda
|
||||||
|
- q-guided-retrieval
|
||||||
benchmarks:
|
benchmarks:
|
||||||
-
|
- LLAB
|
||||||
|
- LiveCodeBench
|
||||||
|
- MMMU-Pro
|
||||||
|
- ERQA
|
||||||
|
- GPQA-Diamond
|
||||||
|
- BFCL
|
||||||
models:
|
models:
|
||||||
-
|
-
|
||||||
datasets:
|
datasets:
|
||||||
- cs.AI
|
- cs.AI
|
||||||
related_concepts:
|
related_concepts:
|
||||||
-
|
- memory-credit-assignment
|
||||||
|
- procedural-memory
|
||||||
related_jobs:
|
related_jobs:
|
||||||
-
|
-
|
||||||
related_experiments:
|
related_experiments:
|
||||||
-
|
- KC-001-agent-memory-pilot
|
||||||
related_projects:
|
related_projects:
|
||||||
-
|
- learning/agent-memory
|
||||||
collection_score: 20
|
collection_score: 20
|
||||||
collection_queries: function-calling
|
collection_queries: function-calling
|
||||||
---
|
---
|
||||||
|
|
||||||
## One-line Takeaway
|
## 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
|
- problem: 单步 memory value update 忽略跨多次任务形成的经验依赖链。
|
||||||
- inferred topics: agent-evaluation, computer-use, embodied-agent, memory, rag, reasoning, tool-use
|
- method: 用 TD(lambda) 沿 provenance DAG 回传 credit，结构深度代替时间步距离。
|
||||||
- arXiv categories: cs.AI
|
- 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。
|
||||||
- collection score: 20
|
- 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?
|
- [Agent Memory evidence matrix](../../learning/agent-memory/evidence-matrix.md)
|
||||||
- Does it include a benchmark, dataset, code, or reproducible experimental setup?
|
|
||||||
- Should it be promoted from `queued` to `skimmed` or `summarized`?
|
|
||||||
|
|
||||||
## Links
|
## Links
|
||||||
|
|
||||||
|
|||||||
+18
-16
@@ -12,48 +12,50 @@ source: arxiv
|
|||||||
collected_at: 2026-07-08
|
collected_at: 2026-07-08
|
||||||
published_at: 2026-05-11
|
published_at: 2026-05-11
|
||||||
updated_at: 2026-05-11
|
updated_at: 2026-05-11
|
||||||
status: queued
|
status: skimmed
|
||||||
relevance: high
|
relevance: high
|
||||||
topics:
|
topics:
|
||||||
- agent-evaluation
|
- agent-evaluation
|
||||||
- memory
|
- memory
|
||||||
- planning
|
- planning
|
||||||
methods:
|
methods:
|
||||||
-
|
- decision-rate-distortion
|
||||||
|
- k-slot-memory
|
||||||
|
- certified-memory-split
|
||||||
benchmarks:
|
benchmarks:
|
||||||
-
|
- LoCoMo
|
||||||
|
- synthetic-decision-diagnostics
|
||||||
models:
|
models:
|
||||||
-
|
-
|
||||||
datasets:
|
datasets:
|
||||||
- cs.AI
|
- cs.AI
|
||||||
related_concepts:
|
related_concepts:
|
||||||
-
|
- decision-sufficient-state
|
||||||
|
- selective-forgetting
|
||||||
related_jobs:
|
related_jobs:
|
||||||
-
|
-
|
||||||
related_experiments:
|
related_experiments:
|
||||||
-
|
- KC-001-agent-memory-pilot
|
||||||
related_projects:
|
related_projects:
|
||||||
-
|
- learning/agent-memory
|
||||||
collection_score: 13
|
collection_score: 13
|
||||||
collection_queries: language-agent
|
collection_queries: language-agent
|
||||||
---
|
---
|
||||||
|
|
||||||
## One-line Takeaway
|
## 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
|
- problem: relevance、salience 和 summary fidelity 没有直接刻画 memory 对未来决策的价值。
|
||||||
- inferred topics: agent-evaluation, memory, planning
|
- method: 在 K 个 runtime state 预算下定义 decision distortion、forgetting boundary，并按 certified conflict 拆分 state。
|
||||||
- arXiv categories: cs.AI
|
- 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%。
|
||||||
- collection score: 13
|
- 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?
|
- [Agent Memory evidence matrix](../../learning/agent-memory/evidence-matrix.md)
|
||||||
- Does it include a benchmark, dataset, code, or reproducible experimental setup?
|
|
||||||
- Should it be promoted from `queued` to `skimmed` or `summarized`?
|
|
||||||
|
|
||||||
## Links
|
## Links
|
||||||
|
|
||||||
|
|||||||
+18
-16
@@ -12,7 +12,7 @@ source: arxiv
|
|||||||
collected_at: 2026-07-08
|
collected_at: 2026-07-08
|
||||||
published_at: 2026-05-14
|
published_at: 2026-05-14
|
||||||
updated_at: 2026-05-14
|
updated_at: 2026-05-14
|
||||||
status: queued
|
status: skimmed
|
||||||
relevance: high
|
relevance: high
|
||||||
topics:
|
topics:
|
||||||
- agent-evaluation
|
- agent-evaluation
|
||||||
@@ -20,42 +20,44 @@ topics:
|
|||||||
- memory
|
- memory
|
||||||
- tool-use
|
- tool-use
|
||||||
methods:
|
methods:
|
||||||
-
|
- cryptographic-provenance
|
||||||
|
- derivation-lineage
|
||||||
|
- sensitive-action-gate
|
||||||
benchmarks:
|
benchmarks:
|
||||||
-
|
- custom-cross-session-memory-harness
|
||||||
|
- AgentDojo-bridge
|
||||||
models:
|
models:
|
||||||
-
|
-
|
||||||
datasets:
|
datasets:
|
||||||
- cs.CR
|
- cs.CR
|
||||||
- cs.AI
|
- cs.AI
|
||||||
related_concepts:
|
related_concepts:
|
||||||
-
|
- memory-provenance
|
||||||
|
- memory-poisoning
|
||||||
related_jobs:
|
related_jobs:
|
||||||
-
|
-
|
||||||
related_experiments:
|
related_experiments:
|
||||||
-
|
- KC-001-agent-memory-pilot
|
||||||
related_projects:
|
related_projects:
|
||||||
-
|
- learning/agent-memory
|
||||||
collection_score: 17
|
collection_score: 17
|
||||||
collection_queries: agent-memory
|
collection_queries: agent-memory
|
||||||
---
|
---
|
||||||
|
|
||||||
## One-line Takeaway
|
## 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
|
- 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。
|
||||||
- inferred topics: agent-evaluation, computer-use, memory, tool-use
|
- method: 为条目附加签名与 derivation DAG，把来源 trust 传播到 action gate。
|
||||||
- arXiv categories: cs.CR, cs.AI
|
- evidence: 自定义 deterministic harness 的三类攻击中，只有 lineage 方案全部达到 0 ASR。
|
||||||
- collection score: 17
|
- 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?
|
- [Agent Memory evidence matrix](../../learning/agent-memory/evidence-matrix.md)
|
||||||
- Does it include a benchmark, dataset, code, or reproducible experimental setup?
|
|
||||||
- Should it be promoted from `queued` to `skimmed` or `summarized`?
|
|
||||||
|
|
||||||
## Links
|
## Links
|
||||||
|
|
||||||
|
|||||||
+20
-16
@@ -12,7 +12,7 @@ source: arxiv
|
|||||||
collected_at: 2026-07-08
|
collected_at: 2026-07-08
|
||||||
published_at: 2026-06-03
|
published_at: 2026-06-03
|
||||||
updated_at: 2026-06-03
|
updated_at: 2026-06-03
|
||||||
status: queued
|
status: skimmed
|
||||||
relevance: high
|
relevance: high
|
||||||
topics:
|
topics:
|
||||||
- agent-evaluation
|
- agent-evaluation
|
||||||
@@ -21,41 +21,45 @@ topics:
|
|||||||
- rag
|
- rag
|
||||||
- tool-use
|
- tool-use
|
||||||
methods:
|
methods:
|
||||||
-
|
- agentic-memory-harness
|
||||||
|
- schema-diagnostics
|
||||||
|
- active-retrieval
|
||||||
benchmarks:
|
benchmarks:
|
||||||
-
|
- LoCoMo
|
||||||
|
- HotpotQA
|
||||||
|
- AMABench
|
||||||
|
- ALFWorld
|
||||||
models:
|
models:
|
||||||
-
|
-
|
||||||
datasets:
|
datasets:
|
||||||
- cs.AI
|
- cs.AI
|
||||||
related_concepts:
|
related_concepts:
|
||||||
-
|
- schema-commitment
|
||||||
|
- agentic-retrieval
|
||||||
related_jobs:
|
related_jobs:
|
||||||
-
|
-
|
||||||
related_experiments:
|
related_experiments:
|
||||||
-
|
- KC-001-agent-memory-pilot
|
||||||
related_projects:
|
related_projects:
|
||||||
-
|
- learning/agent-memory
|
||||||
collection_score: 19
|
collection_score: 19
|
||||||
collection_queries: agent-memory
|
collection_queries: agent-memory
|
||||||
---
|
---
|
||||||
|
|
||||||
## One-line Takeaway
|
## 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
|
- 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。
|
||||||
- inferred topics: agent-evaluation, memory, planning, rag, tool-use
|
- method: 诊断 representation loss 和 retrieval loss，并提出带主动工具循环的 AutoMEM。
|
||||||
- arXiv categories: cs.AI
|
- 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。
|
||||||
- collection score: 19
|
- 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?
|
- [Agent Memory evidence matrix](../../learning/agent-memory/evidence-matrix.md)
|
||||||
- Does it include a benchmark, dataset, code, or reproducible experimental setup?
|
|
||||||
- Should it be promoted from `queued` to `skimmed` or `summarized`?
|
|
||||||
|
|
||||||
## Links
|
## Links
|
||||||
|
|
||||||
|
|||||||
+17
-16
@@ -12,7 +12,7 @@ source: arxiv
|
|||||||
collected_at: 2026-07-08
|
collected_at: 2026-07-08
|
||||||
published_at: 2026-06-04
|
published_at: 2026-06-04
|
||||||
updated_at: 2026-06-04
|
updated_at: 2026-06-04
|
||||||
status: queued
|
status: skimmed
|
||||||
relevance: high
|
relevance: high
|
||||||
topics:
|
topics:
|
||||||
- computer-use
|
- computer-use
|
||||||
@@ -21,41 +21,42 @@ topics:
|
|||||||
- rag
|
- rag
|
||||||
- tool-use
|
- tool-use
|
||||||
methods:
|
methods:
|
||||||
-
|
- hierarchical-execution-state-tree
|
||||||
|
- grow-compress-maintain-revise
|
||||||
|
- error-branch-isolation
|
||||||
benchmarks:
|
benchmarks:
|
||||||
-
|
- MemoryArena
|
||||||
models:
|
models:
|
||||||
-
|
-
|
||||||
datasets:
|
datasets:
|
||||||
- cs.AI
|
- cs.AI
|
||||||
related_concepts:
|
related_concepts:
|
||||||
-
|
- execution-state
|
||||||
|
- long-horizon-agent
|
||||||
related_jobs:
|
related_jobs:
|
||||||
-
|
-
|
||||||
related_experiments:
|
related_experiments:
|
||||||
-
|
- KC-001-agent-memory-pilot
|
||||||
related_projects:
|
related_projects:
|
||||||
-
|
- learning/agent-memory
|
||||||
collection_score: 16
|
collection_score: 16
|
||||||
collection_queries: agent-memory, rag-agent
|
collection_queries: agent-memory, rag-agent
|
||||||
---
|
---
|
||||||
|
|
||||||
## One-line Takeaway
|
## 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
|
- 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。
|
||||||
- inferred topics: computer-use, memory, planning, rag, tool-use
|
- method: 两层 state tree 保存 action-observation 与 subgoal summary，通过 Grow、Compress、Maintain、Revise 管理当前路径。
|
||||||
- arXiv categories: cs.AI
|
- 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。
|
||||||
- collection score: 16
|
- 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?
|
- [Agent Memory evidence matrix](../../learning/agent-memory/evidence-matrix.md)
|
||||||
- Does it include a benchmark, dataset, code, or reproducible experimental setup?
|
|
||||||
- Should it be promoted from `queued` to `skimmed` or `summarized`?
|
|
||||||
|
|
||||||
## Links
|
## Links
|
||||||
|
|
||||||
|
|||||||
+18
-16
@@ -12,7 +12,7 @@ source: arxiv
|
|||||||
collected_at: 2026-07-08
|
collected_at: 2026-07-08
|
||||||
published_at: 2026-06-09
|
published_at: 2026-06-09
|
||||||
updated_at: 2026-06-09
|
updated_at: 2026-06-09
|
||||||
status: queued
|
status: skimmed
|
||||||
relevance: high
|
relevance: high
|
||||||
topics:
|
topics:
|
||||||
- agent-evaluation
|
- agent-evaluation
|
||||||
@@ -20,42 +20,44 @@ topics:
|
|||||||
- rag
|
- rag
|
||||||
- tool-use
|
- tool-use
|
||||||
methods:
|
methods:
|
||||||
-
|
- topic-documents
|
||||||
|
- buffered-consolidation
|
||||||
|
- agentic-retrieval
|
||||||
benchmarks:
|
benchmarks:
|
||||||
-
|
- MemoryAgentBench
|
||||||
|
- LongMemEval
|
||||||
models:
|
models:
|
||||||
-
|
-
|
||||||
datasets:
|
datasets:
|
||||||
- cs.AI
|
- cs.AI
|
||||||
- cs.CL
|
- cs.CL
|
||||||
related_concepts:
|
related_concepts:
|
||||||
-
|
- semantic-memory
|
||||||
|
- selective-forgetting
|
||||||
related_jobs:
|
related_jobs:
|
||||||
-
|
-
|
||||||
related_experiments:
|
related_experiments:
|
||||||
-
|
- KC-001-agent-memory-pilot
|
||||||
related_projects:
|
related_projects:
|
||||||
-
|
- learning/agent-memory
|
||||||
collection_score: 18
|
collection_score: 18
|
||||||
collection_queries: agent-memory
|
collection_queries: agent-memory
|
||||||
---
|
---
|
||||||
|
|
||||||
## One-line Takeaway
|
## 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
|
- problem: 孤立记录难以聚合证据、修订事实和维护跨 session 一致性。
|
||||||
- inferred topics: agent-evaluation, memory, rag, tool-use
|
- method: observation 先进入 buffer，再巩固到 topic documents；读取时迭代 search、grep 和 read_lines。
|
||||||
- arXiv categories: cs.AI, cs.CL
|
- evidence: MemoryAgentBench overall 64.7；LongMemEval 上 summary-only 41.7、summary+BM25 76.0、agentic 79.3。
|
||||||
- collection score: 18
|
- 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?
|
- [Agent Memory evidence matrix](../../learning/agent-memory/evidence-matrix.md)
|
||||||
- Does it include a benchmark, dataset, code, or reproducible experimental setup?
|
|
||||||
- Should it be promoted from `queued` to `skimmed` or `summarized`?
|
|
||||||
|
|
||||||
## Links
|
## Links
|
||||||
|
|
||||||
|
|||||||
+19
-16
@@ -12,7 +12,7 @@ source: arxiv
|
|||||||
collected_at: 2026-07-08
|
collected_at: 2026-07-08
|
||||||
published_at: 2026-06-10
|
published_at: 2026-06-10
|
||||||
updated_at: 2026-06-10
|
updated_at: 2026-06-10
|
||||||
status: queued
|
status: skimmed
|
||||||
relevance: high
|
relevance: high
|
||||||
topics:
|
topics:
|
||||||
- agent-evaluation
|
- agent-evaluation
|
||||||
@@ -23,9 +23,13 @@ topics:
|
|||||||
- rag
|
- rag
|
||||||
- reasoning
|
- reasoning
|
||||||
methods:
|
methods:
|
||||||
-
|
- hierarchical-memory-workspace
|
||||||
|
- memory-skill-evolution
|
||||||
|
- rl-retrieval-agent
|
||||||
benchmarks:
|
benchmarks:
|
||||||
-
|
- ALFWorld
|
||||||
|
- LoCoMo
|
||||||
|
- LongMemEval
|
||||||
models:
|
models:
|
||||||
-
|
-
|
||||||
datasets:
|
datasets:
|
||||||
@@ -33,33 +37,32 @@ datasets:
|
|||||||
- cs.CL
|
- cs.CL
|
||||||
- cs.LG
|
- cs.LG
|
||||||
related_concepts:
|
related_concepts:
|
||||||
-
|
- memory-organization
|
||||||
|
- agentic-retrieval
|
||||||
related_jobs:
|
related_jobs:
|
||||||
-
|
-
|
||||||
related_experiments:
|
related_experiments:
|
||||||
-
|
- KC-001-agent-memory-pilot
|
||||||
related_projects:
|
related_projects:
|
||||||
-
|
- learning/agent-memory
|
||||||
collection_score: 18
|
collection_score: 18
|
||||||
collection_queries: agent-memory
|
collection_queries: agent-memory
|
||||||
---
|
---
|
||||||
|
|
||||||
## One-line Takeaway
|
## 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
|
- 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。
|
||||||
- inferred topics: agent-evaluation, computer-use, embodied-agent, memory, planning, rag, reasoning
|
- 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。
|
||||||
- arXiv categories: cs.AI, cs.CL, cs.LG
|
- 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。
|
||||||
- collection score: 18
|
- 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?
|
- [Agent Memory evidence matrix](../../learning/agent-memory/evidence-matrix.md)
|
||||||
- Does it include a benchmark, dataset, code, or reproducible experimental setup?
|
|
||||||
- Should it be promoted from `queued` to `skimmed` or `summarized`?
|
|
||||||
|
|
||||||
## Links
|
## Links
|
||||||
|
|
||||||
|
|||||||
+17
-16
@@ -12,7 +12,7 @@ source: arxiv
|
|||||||
collected_at: 2026-07-08
|
collected_at: 2026-07-08
|
||||||
published_at: 2026-06-12
|
published_at: 2026-06-12
|
||||||
updated_at: 2026-06-12
|
updated_at: 2026-06-12
|
||||||
status: queued
|
status: skimmed
|
||||||
relevance: high
|
relevance: high
|
||||||
topics:
|
topics:
|
||||||
- agent-evaluation
|
- agent-evaluation
|
||||||
@@ -21,41 +21,42 @@ topics:
|
|||||||
- reasoning
|
- reasoning
|
||||||
- workflow-agent
|
- workflow-agent
|
||||||
methods:
|
methods:
|
||||||
-
|
- budget-matched-baseline
|
||||||
|
- online-augmentation-audit
|
||||||
benchmarks:
|
benchmarks:
|
||||||
-
|
- WebArena
|
||||||
|
- WorkArena-L1
|
||||||
models:
|
models:
|
||||||
-
|
-
|
||||||
datasets:
|
datasets:
|
||||||
- cs.CL
|
- cs.CL
|
||||||
related_concepts:
|
related_concepts:
|
||||||
-
|
- memory-cost-accounting
|
||||||
|
- evaluation-hygiene
|
||||||
related_jobs:
|
related_jobs:
|
||||||
-
|
-
|
||||||
related_experiments:
|
related_experiments:
|
||||||
-
|
- KC-001-agent-memory-pilot
|
||||||
related_projects:
|
related_projects:
|
||||||
-
|
- learning/agent-memory
|
||||||
collection_score: 14
|
collection_score: 14
|
||||||
collection_queries: web-gui-agent
|
collection_queries: web-gui-agent
|
||||||
---
|
---
|
||||||
|
|
||||||
## One-line Takeaway
|
## 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
|
- problem: augmentation 论文常忽略 induction、retrieval、verification 和 prompt injection 的额外 token。
|
||||||
- inferred topics: agent-evaluation, computer-use, memory, reasoning, workflow-agent
|
- 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。
|
||||||
- arXiv categories: cs.CL
|
- 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。
|
||||||
- collection score: 14
|
- 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?
|
- [Agent Memory evidence matrix](../../learning/agent-memory/evidence-matrix.md)
|
||||||
- Does it include a benchmark, dataset, code, or reproducible experimental setup?
|
|
||||||
- Should it be promoted from `queued` to `skimmed` or `summarized`?
|
|
||||||
|
|
||||||
## Links
|
## Links
|
||||||
|
|
||||||
|
|||||||
+19
-16
@@ -12,7 +12,7 @@ source: arxiv
|
|||||||
collected_at: 2026-07-08
|
collected_at: 2026-07-08
|
||||||
published_at: 2026-06-23
|
published_at: 2026-06-23
|
||||||
updated_at: 2026-06-23
|
updated_at: 2026-06-23
|
||||||
status: queued
|
status: skimmed
|
||||||
relevance: high
|
relevance: high
|
||||||
topics:
|
topics:
|
||||||
- agent-evaluation
|
- agent-evaluation
|
||||||
@@ -22,41 +22,44 @@ topics:
|
|||||||
- reasoning
|
- reasoning
|
||||||
- tool-use
|
- tool-use
|
||||||
methods:
|
methods:
|
||||||
-
|
- memory-transition-verification
|
||||||
|
- transition-ranked-grpo
|
||||||
|
- write-revise-prune
|
||||||
benchmarks:
|
benchmarks:
|
||||||
-
|
- MemoryAgentBench
|
||||||
|
- HaluMem
|
||||||
|
- Mem-alpha
|
||||||
models:
|
models:
|
||||||
-
|
-
|
||||||
datasets:
|
datasets:
|
||||||
- cs.AI
|
- cs.AI
|
||||||
related_concepts:
|
related_concepts:
|
||||||
-
|
- memory-consolidation
|
||||||
|
- transition-verification
|
||||||
related_jobs:
|
related_jobs:
|
||||||
-
|
-
|
||||||
related_experiments:
|
related_experiments:
|
||||||
-
|
- KC-001-agent-memory-pilot
|
||||||
related_projects:
|
related_projects:
|
||||||
-
|
- learning/agent-memory
|
||||||
collection_score: 19
|
collection_score: 19
|
||||||
collection_queries: agent-memory
|
collection_queries: agent-memory
|
||||||
---
|
---
|
||||||
|
|
||||||
## One-line Takeaway
|
## 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
|
- 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。
|
||||||
- inferred topics: agent-evaluation, computer-use, memory, rag, reasoning, tool-use
|
- method: Memory Transition Verifier 为局部更新打分，并用 Transition-Ranked GRPO 优化更新策略。
|
||||||
- arXiv categories: cs.AI
|
- 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%。
|
||||||
- collection score: 19
|
- 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?
|
- [Agent Memory evidence matrix](../../learning/agent-memory/evidence-matrix.md)
|
||||||
- Does it include a benchmark, dataset, code, or reproducible experimental setup?
|
|
||||||
- Should it be promoted from `queued` to `skimmed` or `summarized`?
|
|
||||||
|
|
||||||
## Links
|
## Links
|
||||||
|
|
||||||
|
|||||||
Reference in New Issue
Block a user