面向中文技术读者的深度技术笔记，长期研究 AI 记忆系统、AI Native 工作实践与网络安全工程，关注论文、源码、架构方案、验证方法和工程取舍。https://agent-lab.top/https://agent-lab.top/articles/2026-06-15-runtime-memory-poisoning-certified-defense/https://agent-lab.top/articles/2026-06-15-runtime-memory-poisoning-certified-defense/SMSR、MemVenom 和长期记忆安全综述把 Agent 记忆安全推到可验证治理阶段：生产系统不能只做 prompt filter，而要把来源签名、随机化检索、证书复算、回滚和工具调用审计放进同一条验收链。Mon, 15 Jun 2026 00:00:00 GMT安全分析AI memoryagent memorylong-term memorymemory poisoningmemory securityRAGagent securitymemory evaluationhttps://agent-lab.top/articles/2026-06-15-agent-orchestration-cybersecurity-workflows/https://agent-lab.top/articles/2026-06-15-agent-orchestration-cybersecurity-workflows/Agent 编排不是让一个大模型直接接管安全运营，而是把 triage、证据收集、静态分析、威胁情报、检测工程、修复验证和人工审批组织成有状态、有权限边界、可回放的安全工作流。本文给出一套面向 SOC 与白盒扫描的工程方案。Mon, 15 Jun 2026 00:00:00 GMT安全工程agent securitysecurity automationSOCagent orchestrationwhite-box scanningCodeQLSARIFhuman-in-the-loophttps://agent-lab.top/articles/2026-06-14-infini-memory-topic-document-maintenance/https://agent-lab.top/articles/2026-06-14-infini-memory-topic-document-maintenance/Infini Memory 把长期 Agent 记忆从孤立片段和向量召回，推进到可维护的主题文档库。工程上，Topic Document 的价值不只是可读 Markdown，而是把写入缓冲、证据聚合、事实修订、局部检索和审计元数据放进同一个维护单元。Sun, 14 Jun 2026 00:00:00 GMT论文解读AI memoryagent memorylong-term memorymemory-augmented agentsRAGcontext compressionmemory evaluationpersonalizationhttps://agent-lab.top/articles/2026-06-11-memgate-memory-search-trust-boundary/https://agent-lab.top/articles/2026-06-11-memgate-memory-search-trust-boundary/arXiv:2606.06054 MemGate 把个人 Agent 的长期记忆检索定义为信任边界。工程上，记忆读路径不能只按相似度把候选片段塞进上下文，而要在检索和注入之间增加任务条件准入、来源权威、作用域隔离和工具副作用绑定。Thu, 11 Jun 2026 00:00:00 GMT安全分析AI memoryagent memorylong-term memorymemory securityRAGagent securitymemory evaluationpersonalizationhttps://agent-lab.top/articles/2026-06-10-agent-memory-systems-cost-governance/https://agent-lab.top/articles/2026-06-10-agent-memory-systems-cost-governance/arXiv:2606.06448 把 Agent memory 从结构设计推进到系统工作负载表征：写入构建、检索、生成各自承担不同成本。生产系统需要用 phase-aware profiling、容量治理、调度策略和收益指标证明记忆层不是更贵的上下文。Wed, 10 Jun 2026 00:00:00 GMT工程架构AI memoryagent memorylong-term memorymemory evaluationRAGsystemsobservabilitycapacity planninghttps://agent-lab.top/articles/2026-06-08-mpbench-memory-poisoning-write-surface/https://agent-lab.top/articles/2026-06-08-mpbench-memory-poisoning-write-surface/arXiv:2606.04329 把 Agent 记忆投毒从零散案例整理成写入通道、结构性漏洞和 ASR/RSR 评测问题。工程上真正该落地的是记忆写入面的资产清单、来源权威、写后审计和跨会话回归测试。Mon, 08 Jun 2026 00:00:00 GMT安全分析AI memoryagent memorylong-term memorymemory poisoningmemory securityagent securitymemory evaluationprompt injectionhttps://agent-lab.top/articles/2026-06-07-belief-entropy-memory-policy-optimization/https://agent-lab.top/articles/2026-06-07-belief-entropy-memory-policy-optimization/arXiv:2605.30159 提出用 Belief Entropy 给长程 Agent 的递归记忆摘要做中间奖励：问题不只是摘要能否变短，而是每一步摘要后，Agent 对任务状态、缺口信息和后续动作的信念是否更清楚。Sun, 07 Jun 2026 00:00:00 GMT研究分析AI memoryagent memorylong-term memorymemory-augmented agentscontext compressionmemory evaluationRAGreinforcement learninghttps://agent-lab.top/articles/2026-06-06-agentcl-continual-learning-memory-evaluation/https://agent-lab.top/articles/2026-06-06-agentcl-continual-learning-memory-evaluation/从 arXiv:2606.02461 AgentCL 看，Agent 长期记忆评测不应只停留在长对话召回、环境问答或抗干扰测试，还要测前序任务经验能否在后续任务中被稳定复用，以及何时会造成负迁移。Sat, 06 Jun 2026 00:00:00 GMT研究分析AI memoryagent memorylong-term memorycontinual learningmemory evaluationmemory-augmented agentsRAGSWE-Benchhttps://agent-lab.top/articles/2026-06-05-agent-libos-runtime-authority-boundary/https://agent-lab.top/articles/2026-06-05-agent-libos-runtime-authority-boundary/从 arXiv:2606.03895 Agent libOS 看，长期运行 Agent 的风险不只在 prompt、工具描述或扫描规则里，而在调度、对象记忆、权限授予、人类审批、恢复和审计这些运行时原语能否成为真正的授权边界。Fri, 05 Jun 2026 00:00:00 GMT安全工程agent securityruntime securitycapabilityauditMCPwhite-box scanningAI memoryhttps://agent-lab.top/articles/2026-06-04-memguard-type-aware-memory-boundary/https://agent-lab.top/articles/2026-06-04-memguard-type-aware-memory-boundary/从 arXiv:2605.28009 MemGuard 看，长期记忆的可靠性问题不只来自检索召回不足，也来自把稳定事实、情景事件和操作规则混成同一种证据。生产 Agent 记忆层需要类型边界、关系图、查询路由和可审计的组合策略。Thu, 04 Jun 2026 00:00:00 GMT工程架构AI memoryagent memorylong-term memorymemory evaluationmemory reliabilitycontext engineeringRAGhttps://agent-lab.top/articles/2026-06-03-agentir-adaptive-memory-retrieval/https://agent-lab.top/articles/2026-06-03-agentir-adaptive-memory-retrieval/从 arXiv:2605.25092 AgentIR 看，长期对话记忆的读路径不是普通向量检索：索引会持续增长，查询类型会在会话内漂移，dense 通道并不总值得运行。生产 Agent 记忆层应该把检索策略、时间分区、延迟预算、来源治理和写入状态分开设计。Wed, 03 Jun 2026 00:00:00 GMT工程架构AI memoryagent memorylong-term memoryRAGinformation retrievalmemory evaluationLoCoMoLongMemEvalhttps://agent-lab.top/articles/2026-06-02-certified-traces-agent-security-audit/https://agent-lab.top/articles/2026-06-02-certified-traces-agent-security-audit/从 arXiv:2605.24462 的 Certified Traces、AgentSecBench、Agent-BOM 和当前 Agent SDK/Bedrock 工程接口看，安全 Agent 的关键不是让模型解释得更像人，而是让每次工具调用、白盒扫描、修复和部署动作在执行前携带可检查的权限、来源、证据和回放条件。Tue, 02 Jun 2026 00:00:00 GMT安全工程agent securitysecurity audittool usecertified traceswhite-box scannerAgentSecBenchprompt injectionmemory securityhttps://agent-lab.top/articles/2026-06-01-cascading-compaction-parametric-memory/https://agent-lab.top/articles/2026-06-01-cascading-compaction-parametric-memory/arXiv:2605.24657 把软件开发对话里的级联压缩和 LoRA 式权重合并放到同一评测里：压缩循环会快速丢失程序性纠错和项目事实，而权重合并保留更多知识。但这不意味着所有记忆都应该写进模型，真正的问题是如何在上下文、外部记忆和可回滚适配器之间划边界。Mon, 01 Jun 2026 00:00:00 GMT研究分析AI memoryagent memorylong-term memorymemory-augmented agentscontext compressionpersonalizationLoRAmemory evaluationhttps://agent-lab.top/articles/2026-05-31-memory-poisoning-selective-tool-hijack/https://agent-lab.top/articles/2026-05-31-memory-poisoning-selective-tool-hijack/从 MemPoison 和 MemMorph 看，Agent 记忆攻击正在从“把恶意内容写进长期记忆”推进到“让恶意内容通过抽取、重写、检索和工具推理链条”。生产系统需要把记忆写入、来源权威、检索召回和工具授权放进同一套评测。Sun, 31 May 2026 00:00:00 GMT安全分析AI memoryagent memorylong-term memorymemory poisoningmemory-augmented agentsmemory evaluationtool useprompt injectionhttps://agent-lab.top/articles/2026-05-29-memory-injection-working-context/https://agent-lab.top/articles/2026-05-29-memory-injection-working-context/从 SuperBrain、Claude Code hooks、claude-mem 和 Memory-R2 看，长期记忆系统的难点正在从存储迁移到工作上下文注入：什么时候取、取多少、凭什么取、如何阻止旧记忆污染当前任务。Fri, 29 May 2026 00:00:00 GMT工程解读AI memoryagent memorylong-term memorymemory-augmented agentscontext compressionmemory evaluationpersonalizationRAGhttps://agent-lab.top/articles/2026-05-28-persistent-agent-memory-environment-evaluation/https://agent-lab.top/articles/2026-05-28-persistent-agent-memory-environment-evaluation/从 Persistent AI Agents in Academic Research 看，长期记忆 Agent 的评估对象不该只是单轮回答、RAG 命中率或 token 成本，而应扩展到人-代理-文件-工具-计划任务-治理规则组成的持久化环境。Thu, 28 May 2026 00:00:00 GMT论文解读AI memoryagent memorylong-term memorymemory-augmented agentsmemory evaluationcontext compressionpersonalizationRAGhttps://agent-lab.top/articles/2026-05-21-faulty-memory-consolidation-write-path/https://agent-lab.top/articles/2026-05-21-faulty-memory-consolidation-write-path/从 Useful Memories Become Faulty When Continuously Updated by LLMs 看，自动把成功轨迹持续压缩成文字经验，可能让 Agent 从有用记忆退化到错误记忆；长期记忆系统需要把 consolidation 当成可验证、可回滚、可门控的写操作，而不是无条件后台任务。Thu, 21 May 2026 00:00:00 GMT论文解读AI memoryagent memorylong-term memorymemory-augmented agentscontext compressionforgettingmemory evaluationRAGhttps://agent-lab.top/articles/2026-05-20-longmint-memory-interference-agent-evaluation/https://agent-lab.top/articles/2026-05-20-longmint-memory-interference-agent-evaluation/LongMINT 把长期记忆评测推到多目标干扰、事实修订和跨片段聚合推理场景；结合 MedMemoryBench 的 memory saturation，可以看到生产 Agent 记忆的核心风险不是容量不足，而是旧事实、新事实、噪声和多任务目标互相污染。Wed, 20 May 2026 00:00:00 GMT研究综述AI memoryagent memorylong-term memorymemory-augmented agentsmemory evaluationRAGforgettingpersonalizationhttps://agent-lab.top/articles/2026-05-18-agent-cpg-lfp-white-box-scanner/https://agent-lab.top/articles/2026-05-18-agent-cpg-lfp-white-box-scanner/本文把 Agent、Code Property Graph、最小不动点数据流分析、规则引擎和验证沙箱合成一个白盒扫描器方案：不是让大模型直接猜漏洞，而是让它围绕代码图、状态机、证据链和 PoC 验证来工作。Mon, 18 May 2026 00:00:00 GMT网络安全white-box scannerCPGLFPstatic analysisagent securityCodeQLJoernhttps://agent-lab.top/articles/2026-05-18-zipact-state-driven-agent-memory/https://agent-lab.top/articles/2026-05-18-zipact-state-driven-agent-memory/TMLR 2026-05-17 接收的 ZipAct 把 agent 的交互历史压缩成 Goal、World、Constraint 三类结构化状态，让动作生成只看当前状态表和最新观察。它提醒我们，长任务记忆的关键不只是存储更多历史，而是把可执行状态、负反馈和约束持续更新到一个可验证的工作记忆里。Mon, 18 May 2026 00:00:00 GMT论文解读AI memoryagent memorycontext compressionmemory-augmented agentsmemory evaluationRAGforgettinghttps://agent-lab.top/articles/2026-05-15-linkedin-hltm-production-personalization-memory/https://agent-lab.top/articles/2026-05-15-linkedin-hltm-production-personalization-memory/LinkedIn 的 Hierarchical Long-Term Semantic Memory 和 Cognitive Memory Agent 把 agent memory 从“多存一些聊天历史”推进到生产个性化基础设施：schema-aligned 语义树、多视图记忆、身份作用域检索、近线增量更新、可观测来源和端到端质量指标。它的启发是，企业 agent 的长期记忆首先是业务边界、隐私隔离和延迟预算问题，其次才是向量检索问题。Fri, 15 May 2026 00:00:00 GMT研究综述AI memoryagent memorylong-term memorypersonalizationRAGmemory evaluationenterprise AIsemantic memoryhttps://agent-lab.top/articles/2026-05-14-agent-memory-environment-experience/https://agent-lab.top/articles/2026-05-14-agent-memory-environment-experience/LongMemEval-V2 把长期记忆问题从用户聊天历史推进到 web/enterprise agent 的环境经验：静态状态、动态变化、工作流、局部陷阱和前提意识。它提醒我们，生产记忆系统不能只追求 RAG 召回分数，还要证明经验能被压缩、检索、使用，并在延迟成本内帮助 agent 像资深同事一样工作。Thu, 14 May 2026 00:00:00 GMT研究综述AI memoryagent memorylong-term memorymemory evaluationRAGweb agentscontext compressionpersonalizationhttps://agent-lab.top/articles/2026-05-13-agent-memory-runtime-reliability/https://agent-lab.top/articles/2026-05-13-agent-memory-runtime-reliability/OpenAI Agents SDK 的 sandbox memory 文档和 AgentMemory 近期连续修复显示，coding agent 的长期记忆不再只是 RAG 或偏好存储，而是涉及文件化状态、渐进披露、隔离布局、召回正确性、部署持久化、上下文预算和观测面的运行时系统。Wed, 13 May 2026 00:00:00 GMT工程分析AI memoryagent memorylong-term memorycoding agentsRAGcontext compressionmemory evaluationpersonalizationhttps://agent-lab.top/articles/2026-05-12-agent-memory-benchmark-hygiene/https://agent-lab.top/articles/2026-05-12-agent-memory-benchmark-hygiene/PlugMem、gbrain-evals、MemPalace 和 Mem0 等近期材料显示，agent memory 的公开评测正在进入高分密集区；真正重要的问题不再只是 R@5 或 accuracy，而是数据划分、调参污染、成本账本、可复现脚本和生产迁移边界。Tue, 12 May 2026 00:00:00 GMT研究综述AI memoryagent memorylong-term memorymemory evaluationRAGpersonalizationbenchmarkhttps://agent-lab.top/articles/2026-05-10-agent-memory-security-boundary/https://agent-lab.top/articles/2026-05-10-agent-memory-security-boundary/5 月上旬的 Trojan Hippo、MAGE 和 Opal 等研究说明，长期记忆不只是个性化能力，也是跨会话攻击面、隐私泄露面和防护状态本身；生产系统必须把记忆写入、来源、工具权限和遗忘纳入同一个安全模型。Sun, 10 May 2026 00:00:00 GMT安全分析AI memoryagent memorylong-term memorymemory securityprompt injectionpersonalizationmemory evaluationprivacyhttps://agent-lab.top/articles/2026-05-09-database-native-agent-memory-langgraph-mongodb/https://agent-lab.top/articles/2026-05-09-database-native-agent-memory-langgraph-mongodb/MongoDB 在 2026-05-08 为 LangGraph.js 长期记忆发布一等支持，意味着短期 checkpoint、长期 store、语义检索和自动 embedding 正在进入应用数据库；但这解决的是运行时和存储边界，不等于解决记忆写入、作用域、遗忘和个性化误用。Sat, 09 May 2026 00:00:00 GMT工程分析AI memoryagent memorylong-term memoryLangGraphMongoDBRAGmemory evaluationpersonalizationhttps://agent-lab.top/articles/2026-05-07-memagents-memory-bottleneck-diagnostics/https://agent-lab.top/articles/2026-05-07-memagents-memory-bottleneck-diagnostics/从 2026-05-06 的 MemAgents workshop 回顾和 ICLR/OpenReview 论文线索看，AI agent memory 的关键问题正在从“要不要长期记忆”转向写入、压缩、召回、利用和评测瓶颈的可诊断化。Thu, 07 May 2026 00:00:00 GMT研究综述AI memoryagent memorylong-term memorymemory evaluationcontext compressionRAGpersonalizationhttps://agent-lab.top/articles/2026-05-05-memory-dilution-continual-learning-agents/https://agent-lab.top/articles/2026-05-05-memory-dilution-continual-learning-agents/从 arXiv:2604.27003 和 elfmem 看，外部记忆并不会自动解决持续学习；它只是把灾难性遗忘从参数更新搬到检索、表示、评分衰减和经验复用策略里。Tue, 05 May 2026 00:00:00 GMT论文解读AI memoryagent memorylong-term memorycontinual learningmemory evaluationforgettingpersonalizationhttps://agent-lab.top/articles/2026-05-03-contextweaver-dependency-memory-agent/https://agent-lab.top/articles/2026-05-03-contextweaver-dependency-memory-agent/从 arXiv:2604.23069 ContextWeaver 看，工具型 LLM agent 的记忆不只是在历史里检索相似片段，而是要保留当前行动真正依赖的早期证据、决策和执行反馈。Sun, 03 May 2026 00:00:00 GMT论文解读AI memoryagent memorylong-term memorycontext compressionRAGmemory evaluationhttps://agent-lab.top/articles/2026-05-02-memory-scope-contract-agent-memory/https://agent-lab.top/articles/2026-05-02-memory-scope-contract-agent-memory/从 2026-05-02 的 agent memory 产品面讨论、ChatGPT Project-only memory、Claude 个性化功能和 Mem0/Cloudflare 的工程材料看，生产级记忆系统必须先定义用户、项目、任务和运行审计的边界，再谈向量库、图谱和长上下文。Sat, 02 May 2026 00:00:00 GMT架构优化AI memoryagent memorylong-term memorypersonalizationmemory evaluationforgettingRAGhttps://agent-lab.top/articles/2026-05-01-experience-compression-spectrum-agent-memory/https://agent-lab.top/articles/2026-05-01-experience-compression-spectrum-agent-memory/从 arXiv:2604.15877 看，长期记忆、技能和规则不是三个孤立模块，而是同一条经验压缩轴上的不同粒度；真正缺失的是跨层晋升、降级和生命周期治理。Fri, 01 May 2026 00:00:00 GMT论文解读AI memoryagent memorylong-term memorycontext compressionagent skillsmemory evaluationforgettinghttps://agent-lab.top/articles/2026-04-29-hermes-memory-system-openclaw-comparison/https://agent-lab.top/articles/2026-04-29-hermes-memory-system-openclaw-comparison/从 NousResearch/hermes-agent 的官方文档和源码看，Hermes 的记忆系统由小容量常驻记忆、SQLite/FTS5 会话检索、外部记忆提供商和技能系统组成；它和 OpenClaw 的差异不在口号，而在事实源、召回路径、晋升机制和治理边界。Wed, 29 Apr 2026 00:00:00 GMT开源项目分析Hermes AgentOpenClawAI memorylong-term memoryagent memorycontext compressionmemory evaluationhttps://agent-lab.top/articles/2026-04-28-agent-memory-infra-postgres-n8n/https://agent-lab.top/articles/2026-04-28-agent-memory-infra-postgres-n8n/从 2026-04-28 的 agentic-db 发布和 n8n Memori 社区节点看，长期记忆正在向数据库、工作流和托管平台下沉；真正要评估的是写入治理、召回路径、遗忘机制和可观测性。Tue, 28 Apr 2026 00:00:00 GMT工程架构AI memoryagent memoryPostgresn8nRAGmemory evaluationhttps://agent-lab.top/articles/2026-04-28-obsidian-personal-memory-system/https://agent-lab.top/articles/2026-04-28-obsidian-personal-memory-system/系统介绍 Obsidian 的主要能力、实际使用流程、底层技术和 AI 工作流接法：Markdown vault、双向链接、Properties、Search、Canvas、Bases、CodeMirror、插件 API 与同步加密。Tue, 28 Apr 2026 00:00:00 GMT个人知识库ObsidianAI memoryMarkdownPKMknowledge graphCodeMirror