Overcoming LLM Agent Memory Limitations: Datalog-Based 'Lemmalog' Released

When utilizing LLM agents for vulnerability research, problems arise where the model forgets previously established facts or continues with incorrect reasoning during complex investigation processes. Existing memory systems rely on storing and retrieving conversation logs, making it difficult for LLMs to independently recognize when new observations invalidate previous assumptions.

To address this issue, a new approach introducing the concept of Program Analysis has been proposed. Program analysis uses fixed point computation techniques that automatically update only derived results when input facts change. The author applied this principle to LLM memory, building a system that automatically invalidates conclusions dependent on observed facts when those facts change.

Datalog-Based Memory Engine 'Lemmalog'

This idea was materialized in a project called 'Lemmalog', which combines the logic programming language Datalog with LLMs. Lemmalog is designed to focus on defining clear facts and rules, rather than having the LLM directly perform the uncertain task of memory retention.

• Separation of Facts and Rules: Explicitly defines facts such as controls(attacker, object_a) and rules that derive controls_kernel_object based on them.
• Automatic Invalidation: If a fact like points_to(object_a, object_b) is proven incorrect through LLDB debugging or similar methods, the conclusion controls_kernel_object(attacker) derived from this fact is automatically invalidated.
• Shift in LLM's Role: The LLM handles ambiguous parts (e.g., interpreting LLDB output), while the Lemmalog engine manages knowledge by maintaining logical consistency.

This approach reduces the likelihood of errors caused by LLMs reconstructing past information via prompts, and contributes to improving the accuracy and traceability of knowledge in complex multi-step reasoning processes such as vulnerability research.