MEMWM: Memory-Augmented Text-Based World Model
The episode discusses MEMWM, a memory-augmented world model that improves AI agents' planning by storing and retrieving transition rules, state caches, and hard-to-predict facts. Hosts highlight the new Structured State Fidelity metric, which exposes failures hidden by surface metrics, and report gains up to 206% in fidelity and 65% in downstream success across ALFWorld, WebShop, and ScienceWorld, without retraining the policy.