Summary

MemPalace tackles ephemeral AI conversation memory by storing all interactions verbatim and making them searchable. Aimed at AI assistant users, it offers a cost-effective, local-first solution to prevent data loss and enable efficient recall, unlike transient or summary-based systems.

How It Works

The system uses a "Palace" metaphor: conversations are stored raw in ChromaDB and structured into Wings (projects/people), Halls (memory types), Rooms (topics), Closets (pointers), and Drawers (verbatim content). This hierarchy aids retrieval. An experimental, lossy AAAK dialect offers token compression for repeated entities, though it currently regresses benchmark performance and is not the default storage.

Quick Start & Requirements

Install via pip install mempalace. Requires Python 3.9+ and ChromaDB, running locally without internet/API keys post-install. Setup involves mempalace init and mempalace mine. Integration with AI assistants like Claude/ChatGPT uses an MCP server (python -m mempalace.mcp_server), enabling AI-driven recall.

Highlighted Details

• Highest-ever LongMemEval score (96.6% R@5 raw mode) with zero API calls.
• Operates completely offline, free, with no subscriptions or cloud dependencies.
• Palace structure provides a claimed +34% retrieval improvement over basic search.
• Features a local, temporal entity-relationship knowledge graph (SQLite).

Maintenance & Community

Active maintenance is evident, with developers promptly addressing community feedback and bug reports (e.g., AAAK accuracy, integration issues). A Discord community link is provided. Recent updates focus on refining experimental features and fixing reported bugs, indicating responsiveness.

Licensing & Compatibility

Licensed under MIT, allowing broad compatibility and commercial use. Designed for integration with various LLMs, including local models, without data leaving the user's machine.

Limitations & Caveats

The AAAK compression dialect is experimental, lossy, and currently shows benchmark regressions. Features like contradiction detection are not yet fully integrated. Specific issues (macOS ARM64 segfaults, ChromaDB compatibility) are actively being addressed.