MemPalace Is the Open-Source Memory the Buy-Side Has Been Waiting For
Trending hard on GitHub today is MemPalace, an open-source agent memory system that just crossed 53.8k stars. And the timing is almost too perfect. Over the past week we covered Supermemory, Hyper, and Walrus Memory, three commercial bets that agent memory should be something you buy. MemPalace is the open-source counterpunch.
The pitch is local-first. It stores your conversation history as verbatim text and retrieves it with semantic search, no API call required for the core loop, no cloud dependency, embeddings can run on your own machine. The structure is a nice mental model: it organizes everything into wings for people and projects, rooms for topics, drawers for the original content. Under the hood there is a temporal knowledge graph and 29 MCP tools, plus auto-save hooks that quietly capture your Claude Code sessions while you work.
The number that earns it the headline is the benchmark. MemPalace claims 96.6% R@5 on LongMemEval using semantic search alone, 98.4% with hybrid retrieval mixing in keyword and temporal boosting, and north of 99% once you add LLM reranking. It calls itself the best-benchmarked open-source AI memory system, and unlike a lot of repos throwing that phrase around, it actually puts a public eval number behind it.
Here is the real story. Agent memory is now a four-horse race in a single week, and the split is the classic one. Three startups say rent it from us, MemPalace says run it yourself, on your hardware, with the receipts to prove it works. If you are a builder who does not want your agent's entire memory living in someone else's database, this is the one to clone first. https://github.com/MemPalace/mempalace
← Back to all articles
The pitch is local-first. It stores your conversation history as verbatim text and retrieves it with semantic search, no API call required for the core loop, no cloud dependency, embeddings can run on your own machine. The structure is a nice mental model: it organizes everything into wings for people and projects, rooms for topics, drawers for the original content. Under the hood there is a temporal knowledge graph and 29 MCP tools, plus auto-save hooks that quietly capture your Claude Code sessions while you work.
The number that earns it the headline is the benchmark. MemPalace claims 96.6% R@5 on LongMemEval using semantic search alone, 98.4% with hybrid retrieval mixing in keyword and temporal boosting, and north of 99% once you add LLM reranking. It calls itself the best-benchmarked open-source AI memory system, and unlike a lot of repos throwing that phrase around, it actually puts a public eval number behind it.
Here is the real story. Agent memory is now a four-horse race in a single week, and the split is the classic one. Three startups say rent it from us, MemPalace says run it yourself, on your hardware, with the receipts to prove it works. If you are a builder who does not want your agent's entire memory living in someone else's database, this is the one to clone first. https://github.com/MemPalace/mempalace
Comments