Giving AI Agents a Memory That Never Leaves Your Machine
Every agent memory system I looked at wanted the same two things: a database I had to host and an API key I had to pay for. That is a reasonable trade for a product. It is a poor trade for a developer tool that mostly needs to remember what I told it last week.
LongtermMemory-MCP does the same job with neither.
Local All the Way Down
The design goal was that nothing leaves the machine. That shaped every choice:
- Storage is SQLite, running through sql.js compiled to WASM — no server, no connection string, just a file.
- Embeddings come from
all-MiniLM-L6-v2running locally, not from an embeddings API. - Vector search is in-process cosine similarity. At the scale a personal memory store actually reaches, a dedicated vector database is a lot of infrastructure to solve a loop over some floats.
- LLM dependency is none at all.
The project was inspired by mcp-mem0, which does the same thing against PostgreSQL or Supabase with OpenAI embeddings. Setup there is a database plus API keys. Setup here is npx longterm-memory-mcp.
The Tools
The core is what you would expect — save_memory, search_memory, update_memory, delete_memory, memory_stats — with search as the interesting half. Memories carry a type (fact, preference, conversation, task, ephemeral, general), tags, and an importance score, so retrieval can be filtered rather than purely semantic: search_by_type, search_by_tags, search_by_date_range.
That distinction matters more than it sounds. “What did I decide about auth?” is a semantic query. “What are this user’s standing preferences?” is a type filter. Making the agent choose keeps recall precise.
Getting Started
npx longterm-memory-mcpNo database, no keys, no account.
Source and full tool list: github.com/MarcelRoozekrans/LongtermMemory-MCP