this seems similar to the problem they faced: pwning.systems/posts/llm-me...
Their agent needed to course correct a long held belief after realising the belief is incorrect. They ended up carving out “belief” itself into a deterministic system the agent simply uses.
pwning.systems
I accidentally turned LLM memory into program analysis :: pwning.systems
Why I stopped trying to give LLM agents a better memory and instead built Lemmalog, a Datalog engine that maintains an agent's knowledge as analysis state, with provenance, retractions and incremental...
