Memkin is a local-first personal memory system for individual work scenarios. It extracts structured signals from Feishu and AI coding sessions, builds an on-device knowledge graph, and makes it available to any Agent for reading and writing via MCP.
【Problem it solves】 AI Agent sessions have no cross-session memory: every new session requires re-explaining who you are, the project background, and existing decisions. Memkin extracts information scattered across Feishu (direct messages, group chats, email, calendar, documents, tasks) and AI coding sessions (Claude Code, Codex, Hermes) into structured signals—entities, decisions, tasks, findings, knowledge, relationships—stored in a knowledge graph on your own machine, and provides it to any Agent for querying and writing back via MCP. Data is kept entirely on-device throughout.
【Main capabilities】
- Feishu collection: 7 sources in total—direct messages, group chats, email, calendar, documents, tasks, and message search—with incremental collection + historical backfill.
- AI session collection: Claude Code (~/.claude/projects/), Codex (~/.codex/), Hermes/OpenClaw (~/.openclaw/agents/).
- MCP server: 36 tools (15 high-intent tools exposed by default), supporting both stdio and Streamable HTTP transports.
- Signal extraction: an LLM Pipeline extracts 7 categories of structured signals, with two-layer noise filtering (rules + LLM scoring), and every signal is traceable to the original message.
- Hybrid retrieval: tsvector full-text (with Chinese support) + pgvector vectors, with RRF fusion ranking.
- Knowledge graph: signals anchored to entities (people, projects, tools), a directed link graph, and cross-platform person identity merging.
- Privacy: sanitization before writing (reversible / irreversible dual track), zero cloud dependency for storage (PGLite embedded database), optional Ollama local embeddings.
- Resident service: memkin up registers a daemon that starts on boot, performs scheduled collection, and comes with run history and alerts.
- Memory consolidation: hot → warm → cold tiered rotation, dead-link repair, and preference inference.
- Obsidian two-way sync: export to a Markdown vault, and import back after editing.
- Web UI: Dashboard, timeline, force-directed knowledge graph, search.
【How to use】 One-clickInstallation (recommended, register as a background daemon service): the script runs in sequence: install Node runtime (if missing) → npm install -g memkin → open the browser setup wizard (enter LLM API Key) → memkin up registers an auto-start background service on boot, and writes the MCP configuration into AI clients already installed on this machine (Claude Code, Codex, Hermes/OpenClaw). Prerequisites: Node.js >= 18 (the installation script handles this automatically). memkin install writes the MCP configuration and memory usage instructions into local AI clients, supporting Claude Code, Claude Desktop, Cursor, Codex, Windsurf, Hermes/OpenClaw. After connecting, just restart the client. For transport methods (stdio / Streamable HTTP), manual configuration, and hooks instructions, see the MCP Integration Guide.
[Use Cases] In Agents connected to Memkin, you can ask questions directly; answers carry [n] citations and can be traced back to the original messages. The data flow has 5 layers: data source collection → signal extraction → local storage → interface output; the three modules of person identity, memory consolidation, and scheduling cut across all layers.
[Runtime Platforms and Limitations] Runtime platforms: macOS / Linux / Windows (PGLite by default, zero external dependencies). The optional self-managed local Postgres engine supports macOS (arm64 / x64) and Linux (x64 / arm64). You can refer to the configuration instructions (memkin.yaml, ports, authentication).
[Future Directions] More data sources (DingTalk, WeCom, WeChat chat records, local documents); extraction quality improvements (cross-block shared context, weighted admission scoring, entity-based narrative aggregation); Web UI memory editing (currently read-only) and a signal traceability audit view.
[License] Apache License 2.0. For bug reports and feature suggestions, please submit an issue.
