AI Agent Memory MCP Server

What is AI Agent Memory MCP is an agent-agnostic persistent memory layer exposed as an MCP (Model Context Protocol) server. It gives any MCP client - Claude Code, Qoder, Cursor — a shared, project-local memory store that survives across sessions and agents. Memories live in your project's .ai-memory/ directory and travel with the project; they never leave the machine. Three stores linked by memory id: SQLite (structured source of truth + FTS5 keyword search), Chroma (vector store), and a Markdown mirror (human-readable, editable).

How to use

1. Install: pip install ai-agent-memory-mcp
2. On first run a default config is generated at .ai-memory/config.yml. Set an OpenAI-compatible embedding provider (defaults to Volcengine doubao-embedding-vision; SiliconFlow / OpenAI also work) and put the API key in .env.
3. Register with your MCP client. For Claude Code: claude mcp add ai-memory -s user -- python -m ai_memory --agent claude-code --project-from-cwd
4. Call who_am_i() to confirm context, then remember() to store and recall() to retrieve. If you switch embedding models later, rebuild vectors with python tests/rebuild_vectors.py.

Key features

• Local-first: SQLite + Chroma + Markdown mirror, all in .ai-memory/; no cloud, data stays with the project.
• Agent-agnostic: Claude Code / Qoder / Cursor share one store; source_agent stamps each write.
• High-hit-rate retrieval: three-way fusion - vector (0.6) + keyword FTS5 trigram (0.25) + title/tag match (0.15).
• [[link]] cross-references: memories reference each other by [[title]]; recall and get_memory return a links field.
• Any OpenAI-compatible embedding (Volcengine / SiliconFlow / OpenAI / others).
• 8 MCP tools: remember, recall, get_memory, search_memories, update_memory, forget, list_memories, who_am_i.

Use cases

• Give a coding agent durable memory across sessions - past decisions, settled choices, pitfalls hit before.
• Cross-agent handoff: Claude Code, Qoder and Cursor on the same project share one store, with source_agent distinguishing writers.
• A project knowledge base that's both machine-queryable (vector + keyword) and human-readable (Markdown mirror).
• Four categories: user (preferences), project (knowledge), process (work), agent (handoff).

FAQ

• Where are memories stored? - In your project's .ai-memory/ (local SQLite + Chroma + Markdown). Only embedding text goes to your configured provider; memories stay local.
• Is it tied to Claude Code? - No, any MCP client works; it's agent-agnostic (change --agent for Qoder/Cursor).
• Can I switch embedding models? - Yes, any OpenAI-compatible service; edit .ai-memory/config.yml, then rebuild vectors with python tests/rebuild_vectors.py.