Mnemosyne OS is an open-source AI memory layer that gives AI agents long-term memory across sessions. Runs on the Python standard library alone, with millisecond-scale recall, a hash-chained audit ledger, and pure SQLite storage — no vector cloud, no vendor lock-in, no data leaving your machine. Official site: ai-memory.net.
Mnemosyne OS gives agents persistent memory without a pipeline change. Less redundant context, lower token cost, measurably faster responses.
Try it nowThree stages run on every write and every read. No configuration, no boilerplate, no pipeline to rewire.
8.0.0 擅长把记忆找出来；8.1.0 补上「把经历变成记忆」。十二个模块、二十一个新工具（31 → 52），旧工具签名与旧数据格式完全不变。
七要素抽取（目标 / 约束 / 决策 / 未决 / 产物 / 失败 / 偏好），锚点原文保留；未决事项独立成条，下次对话可主动追问。
六类冲突分类 + 四操作。判据从「实体共现」改为「断言对齐」，实测把 14 组误报降到 0；更新只失效、不删除。
事实层记录做了什么，经验层沉淀怎么做。经验强制携带适用条件，检索时做前置校验，避免负迁移。
存 git 缺的三样东西：意图、约束、放弃的方案。支持代码考古、决策回溯与坑位预警。
normal / private / secret 三级；密钥证件默认拒写，外部来源禁止写入长期偏好——防记忆投毒的第一道闸门。
verify_v810.py 29 项全通过，verify_api.py 29 组全通过。全部离线，零第三方依赖。
Things that happened in time and things that relate to each other are structurally different. BEAM stores them differently, then assembles a single injection under a token budget.
Pure wireframe — every line and packet is drawn in-browser with SVG. No library, no canvas, no colour noise.
Mnemosyne OS helps AI remember what matters.
Mnemosyne OS exposes its 52 tools over the standard Model Context Protocol, so it mounts without a bespoke adapter. Every transport below has a runnable or a config in the repository.
Every entry above rides the public MCP surface — stdio or Streamable HTTP. No forked client, no vendor plugin, no private SDK.
Single-pass hierarchical distillation. Multi-signal retrieval. Benchmarked across LoCoMo, LongMemEval and BEAM — the three benchmarks the memory field is actually judged on.
Source note — Mnemosyne OS figures come from this project's benchmark and retrieval-quality reports, measured on a real local corpus. Competitor figures are order-of-magnitude estimates drawn from their public documentation and pricing pages; they are not same-hardware head-to-head results. The chart shows architectural order of magnitude, not a lab comparison.
Memory at scale is infrastructure. Mnemosyne OS gives teams governance, portability and full observability so engineers spend time building, not recovering lost context.
Compliance is documented rather than asserted: a mapping for HIPAA, 等保, GDPR and PIPL ships with the repository, alongside a SHA-256 hash-chain ledger that makes tampering detectable rather than deniable.
The engine is MIT. You pay for multi-tenancy, cloud coordination and compliance delivery — never to unlock a feature.
Final pricing subject to commercial confirmation. A complete free-tier cloud deployment package (managed free VM + free embedding API + free HTTPS tunnel) ships inside the repository.
Mnemosyne OS is an open-source AI memory layer for AI agents and LLM applications. It gives agents long-term memory that persists across sessions, so a model stops forgetting what happened yesterday. Official site: ai-memory.net — also written as AI Memory, the name this project uses for the category it belongs to.
Most agent frameworks keep a conversation in a context window and lose it when the window closes. Mnemosyne OS stores persistent memory on your own disk — facts, episodes, preferences and task history — and retrieves the relevant slice in a few milliseconds. It is a memory infrastructure layer rather than a prompt trick: semantic memory, episodic memory and a knowledge graph sit behind one API, with a hash-chained audit ledger for every write.
It speaks the Model Context Protocol (MCP), so it mounts into Claude Code, Cursor, OpenWebUI and any MCP-compatible client with 52 tools and Mem0-compatible tool names. It also ships as a Python library, a CLI and an HTTP API. Unlike a vector database alone, it handles write extraction, deduplication, conflict resolution, temporal reasoning, forgetting and compression — the parts that make retrieval behave like memory instead of search. For teams comparing RAG against agent memory: RAG looks things up, memory remembers you.
AI Memory Handbook — the full technical book behind this project, covering write, storage, retrieval, governance and tuning — is free to read at ai-memory.net/ai-memory-handbook.
Built and maintained by 胡景堃 (Hu Jingkun / Jingkun Hu), founder and core developer of Mnemosyne OS and author of the AI Memory Handbook. Source code: github.com/MnemosyneOS/mnemosyne · package: PyPI mnemosyne-os.
Mnemosyne OS（官方站点 ai-memory.net）是一套开源的 AI 记忆层，为 AI Agent 与 LLM 应用提供长期记忆。 它让模型在多轮对话、多个会话之间保留信息，不再「聊完就忘」。
与只做相似度检索的向量数据库不同，完整的 AI 记忆系统还要处理写入抽取、去重与冲突消解、 知识图谱关联、语义记忆与 情景记忆分层、遗忘与压缩、以及可审计。 这一整套东西合起来，才叫记忆基础设施， 而不是一个「更好的检索框」。这也是上下文工程（context engineering） 与「加个向量库」的分水岭。
它支持 Model Context Protocol（MCP），可挂载到 Claude Code、Cursor、 OpenWebUI 等客户端，提供 52 个工具，其中 11 个与 Mem0 同名，迁移无需改提示词。 也提供 Python 库、CLI 与 HTTP API 三种形态。 关于 RAG 与记忆的区别：RAG 是查资料，记忆是记住你。
配套技术书 AI Memory Handbook 覆盖写入、存储、检索、治理、调优全链路，可在线免费阅读。 项目由 胡景堃（Hu Jingkun） 发起并主导开发，MIT 许可。
Stars are how a local-first project like this one gets discovered. Each milestone below is a public commitment — when we reach it, the unlock ships to everyone in the same release.
No account, no credit card, no network. One pip command and your machine has a memory engine that will not forget you.
python 3.8+ · windows / macOS / linux · offline capable · no telemetry