Researchers have developed a new world model called MemWM that uses memory to improve predictions of how environment states change in response to agent actions. This is done by conditioning next-state imagination on a curated memory bank of transition rules, state caches, and hard-to-predict facts. The team evaluated the effectiveness of MemWM by comparing it with another method, SFT, and found that it improved factual state preservation by up to 206.3%. They also tested MemWM in planning settings where agents had to select actions based on retrieved task-level skills and step-wise corrective guidance, and found that it outperformed SFT-trained world-model agents by up to a 65.4% relative gain. --- Why it matters: This matters because current world models can be flawed in their predictions, leading to systematic errors. MemWM addresses this issue by incorporating memory into the prediction process, which could lead to more accurate and reliable planning decisions in AI systems. Source: https://arxiv.org/abs/2608.07107