Lemmalog

by Jordy Zomer · Framework · Text & code

Datalog engine storing an agent's investigation state as facts and rules, with retraction, provenance and time-bounded facts; built for vulnerability research

AccessOpen codeLicenseMIT

Evidence

  1. Aug 28
    LongMemEval F1 0.463 (accuracy 0.575) versus SimpleMem 0.480, PropMem 0.550 and GPT-4.1 full context 0.197; about 2,700 tokens per query
  2. Aug 28
    Commenters say stale or disproven facts persisting in LLM memory is a real problem in long Claude research projects; invalidation doesn't propagate