Researcher builds Datalog-based memory system to maintain LLM-derived knowledge in vulnerability research

The author describes using LLM agents for vulnerability research and encountering a recurring problem: after hours of investigation, the model would forget prior findings or fail to update conclusions when assumptions changed. This led to suggestions based on ruled-out approaches or reasoning from invalidated observations.

To address this, the author explored memory systems that store and retrieve observations, but found they did not maintain what is currently known. Instead, the author sought a system that could automatically invalidate conclusions when underlying facts changed.

The solution turned into a Datalog engine for LLMs called Lemmalog. Datalog is a declarative logic programming language where facts and rules derive new facts. For example, facts like 'attacker controls object_a' and 'object_a points to object_b' can derive 'attacker controls a kernel object' via a rule.

When a fact changes—such as discovering that 'object_a does not actually point to object_b'—Lemmalog automatically invalidates any conclusions that depended on that fact. This avoids the need to reprocess entire transcripts and reduces the risk of stale reasoning.

The author notes that removing facts in Datalog is nontrivial when multiple derivations support a conclusion. Lemmalog tracks how facts were derived and updates support when facts change, enabling automatic retractions and provenance queries.

Lemmalog splits responsibilities: the LLM handles unstructured understanding (e.g., interpreting debugger output), while the Datalog engine maintains structured facts and rules. This separation allows the LLM to focus on natural language and code understanding, while the engine handles deterministic inference and updates.

The system also enables asking why a conclusion is true by tracing provenance. For example, a conclusion like 'candidate_3_is_exploitable' can be traced back to observations and rules, making it easier to debug and validate agent reasoning.