Memo
@jagoff
memo is a local-first, long-term semantic memory for AI coding agents (Claude Code, Cursor, Cline, Codex). Runs fully offline with in-process embeddings — MLX on Apple Silicon or CPU sentence-transformers on Linux/Docker — and a sqlite-vec + BM25 hybrid retrieval pipeline with re
概览
What is Memo?
Memo is a local-first, long-term semantic memory server for AI coding agents (Claude Code, Cursor, Cline, Codex). It runs fully offline with in-process embeddings and uses a sqlite-vec + BM25 hybrid retrieval pipeline with reranking.
How to use Memo?
Install via pipx install mlx-memo. Use the memo CLI for command-line interaction, or run the memo-mcp MCP server to connect with compatible AI coding tools.
Key features of Memo
- Local-first and fully offline, no cloud or API keys required.
- Uses MLX (Apple Silicon) or CPU sentence-transformers (Linux/Docker) for embeddings.
- Hybrid retrieval with sqlite-vec + BM25 and reranking.
- Markdown files on disk are the source of truth; SQLite index is rebuildable.
- Recall hook auto-injects relevant memories into every prompt.
- Includes a knowledge graph, temporal/time-machine queries, contradiction detection, and nightly synthesis.
Use cases of Memo
- Provide persistent semantic memory for AI coding agents across sessions.
- Enable offline retrieval of past context, decisions, and code snippets.
- Detect contradictions in project documentation or agent memory.
- Build a searchable knowledge graph from markdown notes.
FAQ from Memo
What makes Memo local-first?
All processing happens on your machine. Embeddings are computed in-process using MLX or sentence-transformers, and the SQLite index is stored locally. No data leaves your environment.
What are the runtime requirements?
On Apple Silicon, Memo uses ML
