GitHub - MemPalace/mempalace: The best-benchmarked open-source AI memory system. And it's free.
MemPalace is an open-source tool that stores and retrieves verbatim conversation history on your machine using semantic search and a pluggable vector-store backend. It supports ChromaDB, SQLite exact vectors, Qdrant, and pgvector, and ships a CLI and Docker images for easy setup without external APIs. Benchmarks show 96.6% recall at top-5 on LongMemEval with zero cloud or LLM calls.
- 96.6% recall at top-5 on LongMemEval with zero cloud or LLM calls, rising to 98.4% with hybrid keyword/temporal boosting and ≥99% with LLM rerank
- Runs fully local by default (verbatim storage, semantic search, ChromaDB/SQLite/Qdrant/pgvector options) so nothing leaves your disk unless you opt into a remote backend
- Also benchmarked at 60.3% R@10 on LoCoMo (88.9% with hybrid v5) and 80.3% R@5 on MemBench
- Ships as CLI, Docker (CPU/GPU) images, and an MCP server for direct integration with tools like Claude Code
MemPalace keeps your chat logs verbatim on your own machine and pulls them back with semantic search—no summaries, no cloud APIs unless you choose an external backend. It organizes data into a palace metaphor: “wings” for people or projects, “rooms” for topics, and “drawers” for the original text. The retrieval layer is pluggable, defaulting to ChromaDB but allowing sqlite_exact for local exact checks or optional Qdrant and pgvector backends. You opt in if you want remote storage; nothing leaves your disk by default.
Installation offers three paths. The recommended route uses uv or pipx to sandbox dependencies:
• uv tool install mempalace; mempalace init ~/projects/myapp
• pipx install mempalace
For environments without Python, there’s a Docker image that persists under /data. You can build a CPU image or a GPU variant with CUDA support. Both come bundled with extract and spellcheck extras unless you customize at build time.
CLI commands cover mining and search. Run mempalace mine on a project directory or Claude Code sessions, then mempalace search "your query" to retrieve top hits. You can load context for a new model session via mempalace wake-up. To connect MemPalace as an MCP (Model Communication Protocol) server—say for Claude Code—you define it in your client’s JSON config and point to the Docker command that launches the stdio-based server.
Benchmarks show 96.6% recall at top-5 on LongMemEval (500 questions) without any heuristics or LLM. An optional hybrid pipeline pushes recall to 98.4% using keyword boosting and temporal proximity. Add an LLM rerank to hit ≥99%. Other benchmarks include 60.3% R@10 on LoCoMo (session-level) rising to 88.9% with hybrid v5, and 80.3% R@5 on MemBench. All results and reproduction steps live under benchmarks in the GitHub repo.