Memento 3 Enables LLM World Model Learning via Reflective Rulebooks

Memento 3 allows frozen LLM agents to learn explicit, revisable world models using a natural-language rulebook as external memory. The system compiles these rules into executable code, using prediction errors to iteratively refine hypotheses about environment dynamics.