Milla Jovovich & Ben Sigman · 2026-04-06 · notable
MemPalace — Local AI Memory System (Method of Loci)
Free MIT memory layer for LLMs that stores conversations verbatim, organizes them into a hierarchical palace metaphor, and recalls context at 170-token startup cost. Scored 96.6% on LongMemEval. 44.5k stars.
Store AI conversations verbatim locally using a method-of-loci hierarchy; retrieve them with 96.6% accuracy at a 170-token startup cost.
Key specs
| License | MIT |
|---|---|
| GitHub stars | 44.5k |
| Long mem eval r@5 | 96.6% |
| Startup tokens | 170 |
What is it?
MemPalace is a free, MIT-licensed persistent memory layer for LLMs and AI agents. Instead of summarizing past sessions (which loses details and reasoning), it stores every conversation verbatim in a local ChromaDB vector store and organizes it using a spatial hierarchy borrowed from the ancient method of loci: Wings for people or projects, Rooms for topics, Halls for memory types. It ships with a 19-tool MCP server and works with Claude Code, ChatGPT, Cursor, and Gemini.
How does it work?
Writes use deterministic extraction with no LLM involvement: conversations are chunked into 800-character segments, embedded via a local model, and stored in ChromaDB and SQLite. Reads run a two-pass retrieval — first classify the query into a Hall for high-precision results, then search the full corpus with hall-based reranking. The system boots in roughly 170 tokens by loading only identity context and a compressed top-15 memories snapshot, then fetches deeper context on demand.
Why does it matter?
Most AI memory tools either summarize aggressively (losing context) or require cloud API calls (adding cost and latency). MemPalace's verbatim local approach scored 96.6% on LongMemEval while running entirely on-device. The viral launch — 44.5k GitHub stars in under a week — reflects real demand for a local-first alternative to hosted memory services like Mem0 and Zep.
Who is it for?
Developers adding persistent memory to local AI agents or coding assistants.