MemPalace

MemPalace is a free, open-source AI memory system that stores every conversation, decision, and debugging session verbatim in a local ChromaDB instance — nothing is summarized or discarded. It achieves the highest published LongMemEval R@5 score (96.6% in raw mode, 100% with optional reranking) without requiring any API key or cloud service. The system organizes memories into a hierarchical "palace" structure — wings, halls, rooms, closets, and drawers — inspired by the ancient method of loci, giving AI agents a navigable map rather than a flat search index. It integrates with Claude Code, Claude, ChatGPT, Cursor, Gemini, and local models via an MCP server exposing 19 tools.

• Palace structure — Memories are organized into wings (people/projects), halls (memory types), and rooms (specific topics), delivering a +34% retrieval improvement over unfiltered search.
• Raw verbatim storage — All content is stored as-is in ChromaDB without LLM summarization, which is the source of the 96.6% benchmark score; run pip install mempalace and mempalace init <dir> to get started.
• MCP server with 19 tools — Connect once via claude mcp add mempalace -- python -m mempalace.mcp_server and your AI automatically calls mempalace_search, mempalace_kg_query, and other tools on demand.
• Temporal knowledge graph — SQLite-backed entity-relationship triples with validity windows let you query what was true at any point in time, similar to Zep's Graphiti but fully local and free.
• AAAK compression dialect (experimental) — A lossy abbreviation system for packing repeated entities into fewer tokens at scale; currently regresses LongMemEval vs raw mode (84.2% vs 96.6%) and is not the storage default.
• Specialist agents — Define focused agents (reviewer, architect, ops) each with their own wing and AAAK diary in the palace; agents persist expertise across sessions without bloating your config.
• Auto-save hooks — Claude Code hooks trigger structured saves every 15 messages and before context compaction, keeping the memory palace current automatically.
• Three mining modes — projects for code and docs, convos for chat exports (Claude, ChatGPT, Slack), and general for auto-classification into decisions, preferences, milestones, problems, and emotional context.
• Fully local — Requires only Python 3.9+, ChromaDB, and PyYAML; no internet connection after install, no data leaves your machine.