Agent Memory 每日论文综述
本报告自动生成自 papers.cool/arxiv/cs.AI
筛选标准：标题或摘要包含 agent、memory、RAG、episodic memory 等关键词
生成时间：2026/4/29 20:44:00
📊 今日概况
- 总扫描论文: 25 篇
- Agent Memory 相关: 9 篇
📝 相关论文列表
1. Recursive Multi-Agent Systems
arXiv ID: 2604.25917
核心要点: recursivemas,recursive,agent,latent,collaboration,recursion,multi,computation,end,looped…
关键词: recursivemas,recursive,agent,latent,collaboration,recursion,multi,computation,end,looped
2. ADEMA: A Knowledge-State Orchestration Architecture for Long-Horizon Knowledge Synthesis with LLMAgents
arXiv ID: 2604.25849
核心要点: artifact,knowledge,orchestration,adema,resume,architecture,commitments,interruption,checkpoint,horizon…
关键词: artifact,knowledge,orchestration,adema,resume,architecture,commitments,interruption,checkpoint,horizon
3. StratFormer: Adaptive Opponent Modeling and Exploitation in Imperfect-Information Games
arXiv ID: 2604.25796
核心要点: opponent,stratformer,exploitation,gto,exploitability,opponents,imperfect,agent,per,games…
关键词: opponent,stratformer,exploitation,gto,exploitability,opponents,imperfect,agent,per,games
4. Scalable Inference Architectures for Compound AI Systems: A Production Deployment Study
arXiv ID: 2604.25724
核心要点: compound,invocations,agentic,inference,production,latency,enterprise,deployment,workloads,systems…
关键词: compound,invocations,agentic,inference,production,latency,enterprise,deployment,workloads,systems
5. RADD: Retrieval-Augmented Discrete Diffusion for Multi-Modal Knowledge Graph Completion
arXiv ID: 2604.25693
核心要点: radd,mmkgc,retriever,denoiser,shortlist,kge,reranking,retrieval,completion,discrete…
关键词: radd,mmkgc,retriever,denoiser,shortlist,kge,reranking,retrieval,completion,discrete
6. Think Before You Act – A Neurocognitive Governance Model for Autonomous AI Agents
arXiv ID: 2604.25684
核心要点: governance,neurocognitive,agents,internalized,compliance,agent,humans,think,autonomous,enterprise…
关键词: governance,neurocognitive,agents,internalized,compliance,agent,humans,think,autonomous,enterprise
7. OxyGent: Making Multi-Agent Systems Modular, Observable, and Evolvable via Oxy Abstraction
arXiv ID: 2604.25602
核心要点: oxygent,oxy,evolvable,mas,abstraction,observability,agent,modular,observable,evolution…
关键词: oxygent,oxy,evolvable,mas,abstraction,observability,agent,modular,observable,evolution
8. Plausible but Wrong: A case study on Agentic Failures in Astrophysical Workflows
arXiv ID: 2604.25345
核心要点: agentic,workflows,failure,failures,silent,plausible,astrophysical,scientific,incorrect,wrong…
关键词: agentic,workflows,failure,failures,silent,plausible,astrophysical,scientific,incorrect,wrong
9. AutoResearchBench: Benchmarking AI Agents on Complex Scientific Literature Discovery
arXiv ID: 2604.25256
核心要点: autoresearchbench,research,scientific,agentic,literature,browsing,agents,autonomous,discovery,papers…
关键词: autoresearchbench,research,scientific,agentic,literature,browsing,agents,autonomous,discovery,papers
研究趋势
今日找到 9 篇 Agent Memory 相关论文。从标题和摘要看，主要集中在记忆架构、检索机制和长期记忆存储等方面。
技术演进
当前 Agent Memory 系统正在从简单的 RAG 向更复杂的记忆架构演进，包括分层记忆、情境感知检索和记忆压缩等技术。
关键洞察
- Memory 正成为基础设施: 新 Agent 项目应优先设计 Memory System
- 检索机制持续优化: 从简单向量检索到混合检索
- 记忆压缩受到关注: 避免记忆无限增长
- 多模态记忆: 从文本到图像、视频的统一记忆
开源项目关联
这些研究与 LangChain、LlamaIndex 的记忆模块设计理念一致，可以考虑借鉴到 MyClaw 项目中。
下一步行动
- 研究 Mem0 的记忆管理机制
- 评估向量数据库选型（Qdrant、Weaviate、Chroma）
- 设计 MyClaw 的分层记忆架构
注：GLM-5 API 调用失败，此为备用分析
📚 附录
搜索关键词
agent, memory, memory-augmented, episodic, long-term, recall, retrieval, knowledge base, RAG, retrieval-augmented, episodic memory, working memory, memory system, remember, experience replay, memory network, external memory, vector database
本报告由 OpenClaw 自动生成（GLM-5 深度分析版）
面向 Agent Memory 系统设计者，提供前沿研究洞察