MemPalace releases local-first AI memory system with 96.6% retrieval recall and no API calls required

MemPalace stores conversation history as verbatim text with semantic search retrieval, structured by people/projects (wings), topics (rooms), and original content (drawers), with a pluggable backend defaulting to ChromaDB.

Core benchmark achieves 96.6% R@5 (recall at top 5 results) on LongMemEval with 500 questions using raw semantic search—requiring no API key, cloud service, or LLM. Hybrid v4 with tuning reaches 98.4% R@10 on held-out 450 questions; LLM reranking reaches ≥99% but requires a capable model.

Nothing leaves your machine unless you opt in. System runs locally with ~300 MB disk for the default embedding model and requires Python 3.9+. Installable via `pip install mempalace` with MCP server support, knowledge graph backed by SQLite, and auto-save hooks for Claude Code.