Back to Repositories

milla-jovovich/mempalace

MemPalace is the highest-scoring free AI memory system ever benchmarked. 96.6% on LongMemEval R@5 in raw mode. 100% with Haiku reranking. Zero API costs. Runs entirely on your machine.The core problem it solves: AI amnesia. Every time you start a new chat, your AI assistant forgets everything from the last session. MemPalace gives LLMs persistent, cross-session memory by storing conversations verbatim and making them retrievable through semantic search.The architecture uses a spatial metaphor inspired by the ancient memory palace mnemonic technique. Memories are organized into wings (people or projects), halls (types of memory), rooms (specific ideas), tunnels (cross-wing connections), and drawers (individual entries). This structure delivers a reported 34% retrieval improvement over flat ChromaDB semantic search alone.Technically, it runs on ChromaDB for vector storage and SQLite for the knowledge graph. No cloud services. No API keys. Everything stays local. That matters for privacy-sensitive use cases and for anyone who does not want a $19-$249/month Mem0 bill or a $25+/month Zep subscription.The Benchmark ControversyTransparency matters here. MemPalace initially claimed 100% on LongMemEval, which the community challenged. The team revised it to 96.6%, acknowledging the original score reflected overfitting to test questions. The LoCoMo 100% score is also less impressive than it sounds because retrieval windows exceeded the candidate pool size. Credit to the creators for revising publicly, but be aware that real-world accuracy may differ. One developer reported a 17% correct answer rate when integrated with LLMs in a production workflow.Who Built ThisCo-created by actress Milla Jovovich and crypto CEO Ben Sigman using Claude Code. Jovovich started using AI intensively in late 2025, hit the amnesia problem, and built a solution. The project went viral after its April 6, 2026 launch, accumulating 44,800+ stars and 5,800+ forks.Supported ClientsWorks with Claude Code (native plugin marketplace), Claude Desktop, ChatGPT, Cursor, Gemini, and local models like Llama and Mistral via CLI. Exposes 29 MCP tools for integration.LimitationsProject is 8 days old. Features advertised in the README (like contradiction detection) are missing from the actual codeMCP stdout corruption bug breaks Claude Desktop integration (GitHub Issue #225)No team collaboration features. Solo developer tool onlyAAAK compression mode regresses performance versus raw verbatim storageFor more AI memory and developer tools, browse repos.skila.ai. For commercial AI memory solutions, check tools.skila.ai. For related articles on AI development workflows, visit news.skila.ai.

other
Python

Why It Matters

MemPalace challenges the assumption that AI memory needs expensive LLM summarization. Its 96.6% raw accuracy with zero API calls makes persistent memory accessible to indie developers and small teams who can't afford cloud memory services like Mem0 ($19-249/month) or Zep ($25/month+). The verbatim storage approach — keeping everything instead of letting AI decide what matters — is a philosophical bet that storage is cheaper than intelligence. At 38.8K stars in under a week, the developer community clearly agrees.

Repository Stats

Stars
44.8k
Forks
5.8k
Last Commit
4/10/2026

Category

Related Resources

Weekly AI Digest