Memento 3 is a model-based recursive self-improvement system for frozen LLM agents. It enables agents to learn explicit world models through external memory, allowing them to revise hypotheses about environment dynamics and improve their performance.