Together with engineer Ben Sigman, they released as open source the project MemPalace, which scored 100% on the LongMemEval benchmark. For comparison, popular commercial tools like Mem0 are nervously smoking on the sidelines with their 30-45%.
As Milla herself writes on Instagram*, flat vector search across thousands of text chunks turns the vector DB into a "warehouse filled with junk" where finding the right piece becomes impossible.
🏰 MemPalace uses the principle of the "Memory Palace" (mnemonic technique of the ancient Greeks).
Instead of forcing the neural network to compress text and hallucinate, the system retains all verbatim text (direct speech), but rigidly structures it at the metadata level in ChromaDB:
1️⃣
Wing — entity (specific project or person).2️⃣
Hall — memory type (facts, events, preferences).3️⃣
Room — specific topic (e.g., auth-migration).4️⃣
Drawer — original text chunk.When the agent searches for info, the system does not perform a blind similarity search across the entire database. It filters: Wing -> Hall -> Room. This hierarchy alone increases recall by 34%.
What's under the hood:
▫️ The basic version achieves 96.6% recall on LongMemEval without any LLM API calls (pure Python + ChromaDB + local embeddings). Zero cost.
▫️ Hybrid mode (vector search + LLM reranker on cheap Claude Haiku) pushes the metric to 100%.
▫️ No heavy graph databases. The Temporal Knowledge Graph is implemented on regular local SQLite.
▫️ Built-in compressed dialect
AAAK. This is a symbolic language for agents that compresses context by 30x without loss of meaning, to avoid eating up the token window when loading the agent.Ben Sigman on his X* justifiably proud of the results and invites forking the repository.
🔗 Code is here: MemPalace
Standard note for the authorities:
Instagram is owned by Meta, which is recognized as an extremist organization and banned in the Russian Federation.
The social network X is blocked in the Russian Federation.
#good_opensource
Comments
0No comments yet.
Sign in to join the discussion.