Memkin

What AndreLYL/memkin MCP server does

AndreLYL/memkin MCP server turns information from supported workplace and coding sources into searchable personal memory. It can ingest Feishu/Lark private messages, group chats, email, calendars, documents, tasks, and message search, along with sessions from Claude Code, Codex, and Hermes/OpenClaw. Extracted material is represented as entities, decisions, tasks, discoveries, knowledge, and relationships, then associated with people, projects, and tools.

The resulting store is a private knowledge graph kept on the user’s machine. Agents can query the graph and write memory back through MCP. Responses can include references to the source messages, allowing an agent or user to trace an answer back to collected material.

How it works

Collection supports incremental updates and historical backfills. Content passes through chunking, rule-based and LLM-based noise filtering, structured signal extraction, scoring, and privacy redaction before storage. Signals retain links to their original messages, while identity resolution can associate the same person across platforms.

Search combines Chinese-capable PostgreSQL full-text search with vector retrieval and reciprocal-rank fusion. The default embedded database is PGLite, so the basic storage path does not require an external database. An optional self-managed local PostgreSQL engine and optional Ollama embeddings are also documented.

The MCP interface supports stdio and Streamable HTTP transports. The project describes 36 available tools, with 15 high-intent tools exposed by default. Examples include recall for combining conversations and follow-up tasks, daily_report for aggregating daily activity, prep_for_person for preparing around an interaction, troubleshoot for ordered diagnostic guidance, and get_session_context for project decisions, tasks, and recent timeline information.

Setup and configuration

AndreLYL/memkin MCP server runs on macOS, Linux, and Windows, with Node.js 18 or newer required. The quick trial command is npx memkin start; when configuration is missing, it opens a setup flow and then starts the service and Web UI. A separate installation script can install Node when needed, install the npm package globally, launch setup, register the background service, and write MCP configuration for detected clients.

The memkin install command can configure supported clients, including Claude Code, Claude Desktop, Cursor, Codex, Windsurf, and Hermes/OpenClaw. It also supports a dry-run mode and selecting a specific agent. Historical Claude Code data can be extracted with memkin extract --source claude-code, and optional Claude Code hooks can provide automatic recall. Configuration includes an LLM API key; the exact provider and environment variable names are not specified in the supplied material.

Tools and capabilities

• Collect Feishu/Lark sources incrementally or through historical backfill.
• Import Claude Code, Codex, and Hermes/OpenClaw sessions.
• Search memory using combined full-text and vector retrieval.
• Maintain entities, relationships, decisions, tasks, and source references.
• Run a background daemon with scheduled collection, run history, and alerts.
• Export to an Obsidian Markdown vault and import edits back into memory.
• Inspect data through a dashboard, timeline, graph view, search interface, CLI, REST API, or MCP.

Limitations and notes

The roadmap lists additional sources such as DingTalk, WeCom, WeChat chat records, and local documents as future work, so they should not be treated as currently supported inputs. Web UI memory editing is also listed as planned; the current interface is described as read-only for editing purposes. The system can use an LLM API key for extraction, while Ollama embeddings are optional. Although MCP supports read and write operations, the supplied material does not specify the complete individual tool schema or all configuration fields.