Continual World Modeling via Sleep

Long-lived agents need predictive knowledge that remains useful as experience accumulates. We study continual learning of action-conditioned world models, which receive supervision from both successful and failed interactions and support control through a fixed planner. We introduce Sleep-CWM, which alternates wake, when low-rank adapters learn from incoming experience, with sleep, when the base model consolidates new and replayed experience. Sleep protects event and precondition windows from eviction and allocates replay according to measured forgetting; a generative variant reconstructs replay targets from stored cues. We evaluate a 0.7B video diffusion world model in Minecraft, separating denoising-based prediction scores from sampled futures and planner behavior. In an offline six-stage stream with a ten-window routine buffer, Sleep-CWM achieves 97% success on the hidden-pit task, compared with 0% for uniform replay. Protected windows use additional storage beyond that routine budget. With a larger buffer, uniform replay retains a positive prediction score yet still fails at planning. Sleep-CWM preserves earlier hazard warnings and behavior in an unseen lane, while prediction-level gains in transfer, generative replay, and online learning do not consistently improve behavior. Continual world models therefore require evaluation through sampled futures. Project website is \href{https://yewr.github.io/cwm}{here}.