Lemmalog turns agent memory into a Datalog engine and beats full-transcript prompting on 38x less context.

MIT-licensed Rust, created August 27. It stores LLM observations as provenance-tracked facts with stratified rules and incremental derivation, exposed to a harness over MCP, so the model does language extraction while the database keeps logical consistency. Reported: 0.463 ±0.010 F1 on LongMemEval at 38x smaller context than full-transcript prompting, and 0.533 ±0.001 F1 on LoCoMo at 6x smaller, strongest on knowledge updates and adversarial questions. Splitting extraction from consistency is the right decomposition, and Datalog is a defensible choice for the second half. (GitHub)