MemWM Enhances World Models with Memory for Factual Accuracy

World models are crucial for AI agents, enabling them to plan by predicting how environments evolve based on actions. However, even fluent predictions can suffer from systematic errors, such as omitting critical facts, corrupting attributes, or applying incorrect transition rules. These inaccuracies can severely impact an agent's planning capabilities.To address these limitations, researchers have developed MemWM, a memory-augmented text-based world model. MemWM incorporates a "world memory," which is a carefully curated bank containing transition rules, state caches, and facts that are typically difficult for models to predict. This memory is used to condition the model's next-state imagination, ensuring greater factual accuracy.The effectiveness of MemWM was evaluated using Structured State Fidelity (SSF), a metric that scores predicted states based on benchmark-specific facts. Memory-augmented training improved SSF by over 200% compared to standard supervised fine-tuning. When integrated into a full planning setting, with the policy model frozen, MemWM-augmented agents achieved up to a 65.4% relative gain in downstream success across various benchmarks like ALFWorld, WebShop, and ScienceWorld. This demonstrates that memory-augmented world models can significantly enhance an agent's ability to plan and succeed in complex environments.