ARROW: Augmented Replay for RObust World models

In three linesARROW extends DreamerV3 with a dual-buffer replay system (short and long-term) for continual reinforcement learning. Tested on Atari and Procgen, the algorithm reduces catastrophic forgetting while maintaining knowledge transfer, inspired by experience replay in the brain.

Summary generated by Claude — human-verified