milla-jovovich/mempalace

MemPalace is a free, open-source AI memory system that hit 38,800 GitHub stars and 4,900 forks within days of launching on April 6, 2026. Created by actress Milla Jovovich and engineer Ben Sigman, it holds the highest LongMemEval benchmark score of any free memory framework: 96.6% in raw verbatim mode across 500 questions with zero API calls. The core design is radical in its simplicity. Instead of using an LLM to summarize or extract memories, MemPalace stores complete conversation text verbatim in ChromaDB. The "Palace Architecture" organizes memories through wings (projects/people), rooms (topics), halls (memory types), closets (summaries), and drawers (original files). This hierarchical structure delivers 34% better retrieval versus flat vector search. Token efficiency is the headline feature. MemPalace needs just 170 tokens for startup recall. Estimated annual cost: roughly $10 with search queries — versus $507/year for LLM-based summarization approaches. No API key required at the raw storage level. The hybrid mode (v4 + Haiku reranking) achieves a perfect 100% on LongMemEval — 500/500 questions, all 6 question types at 100%. But this requires an LLM call for reranking, which adds cost and latency. Worth noting: the project caught controversy. The initial release overclaimed features, including inflated compression stats and token count errors. The team published a transparency note in April 2026 acknowledging these issues and correcting the README. The experimental AAAK compression layer scores 84.2% — well below the raw mode — and the team now admits it "trades fidelity for token density." Integrates with Claude Code (marketplace plugin), MCP servers for Claude/ChatGPT/Cursor/Gemini, and Python API for local models like Llama and Mistral. MIT licensed, Python 3.9+. If you're building AI agents that need persistent memory, compare MemPalace with tools in our AI tools directory or explore other trending AI repositories.