I accidentally turned LLM memory into program analysis

A researcher discovered that building memory systems for LLM agents used in vulnerability research is essentially a program analysis problem. When investigating code, LLMs struggle to maintain consistency across long sessions, forgetting or contradicting previously established facts, which mirrors the challenge of tracking logical dependencies in program analysis. The solution involves treating observations and conclusions as facts and rules in a formal system, rather than simply storing and retrieving conversation snippets.