I accidentally turned LLM memory into program analysis

The author describes how LLM agents lose track of established facts during lengthy vulnerability research sessions, prompting a shift from conventional memory retrieval to a program-analysis approach. By treating observations as facts and applying Datalog-style rules, they built a system that automatically invalidates derived conclusions when underlying facts change, enabling more reliable, incremental reasoning.