memory-mcp
Enables personal cross-tool memory by exposing a searchable episodic archive of AI sessions and a curated, high-trust semantic facts layer through MCP tools and resources.
README
memory-mcp
Personal cross-tool memory: a searchable episodic archive of AI sessions from all devices plus a curated, high-trust semantic facts layer, exposed through an MCP server. Designed for a self-hosted arm64 box, WG-only. See docs/mcp/memory-mcp-spec.md for the full design.
Layout
| Package | Purpose |
|---|---|
packages/core |
memory-core library — stores, chunking, embedding, hybrid retrieval, facts, scrub, distill (no MCP, no HTTP) |
packages/cli |
memoryctl — init, verify, ingest, distill, review, stats, export-vault, reembed |
packages/server |
memory-mcp MCP server (Streamable HTTP, static bearer) |
packages/evals |
retrieval golden-set harness (recall@5, MRR, p95) |
deploy/ |
compose.prod.yaml, Dockerfile, systemd timers, backup script |
Two layers, two trust levels
- Episodic archive (Qdrant
memory_archive): every chunk of every session. Dense (bge-m3, 1024) + sparse (IDF-modified TF) vectors, RRF-fused hybrid search. Searched only on demand (search_archive). - Semantic facts (Postgres): curated statements with provenance. Default
read surface (
search_memory). Written byremember(sourceuser) and by human-approved distillation (sourcedistilled) — nothing distilled becomes active withoutmemoryctl review.
MCP surface
Tools: remember, search_memory, search_archive, forget (destructive,
two-phase by query). Resources: memory://stats, memory://facts/recent,
memory://facts/{id}.
Development
pnpm install
pnpm build # all packages
pnpm test # core + server suites (stores are mocked; nothing external touched)
pnpm eval # golden-set retrieval harness (needs live stores + ingested data)
Config is environment-driven through packages/core/src/config.ts
(loadConfig) — see deploy/.env.example for every variable.
Operations
cd deploy && cp .env.example .env # fill in POSTGRES_PASSWORD, STATIC_BEARER, keys
docker compose -f compose.prod.yaml up -d --build
docker compose -f compose.prod.yaml exec memory-mcp memoryctl init
docker compose -f compose.prod.yaml exec memory-mcp memoryctl verify
compose.prod.yaml assumes Qdrant and an OpenAI-compatible LLM gateway
already run as external containers on your host — edit the networks:
section's name: fields to match your own Docker network names.
Daily ingest+distill and nightly backups are systemd timers
(deploy/systemd/). Backups age out after 14 days so a forget propagates
out of backups within that window.
Security posture
- WG-only bind, static bearer (v1); gateway JWT is a planned future auth mode (Step 8), pending a compatible OIDC/JWT gateway in front of the server.
- Scrubbing (gitleaks + custom regexes) is fail-closed and runs before anything is embedded or stored. Unparseable input quarantines the whole file (recorded in the ledger); transient infrastructure failures (embedder, Qdrant, gitleaks unavailable) roll back cleanly, record nothing, and the file is retried on the next run.
forgetis a hard delete in both stores.- Archive chunks are returned inside untrusted-data delimiters — historical transcripts can contain instruction-like text and must be treated as data.
See SECURITY.md to report a vulnerability.
Contributing
Issues and PRs are welcome — see CONTRIBUTING.md for dev setup and project conventions, and CHANGELOG.md for history. This project follows the Code of Conduct.
License
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。