The standard account of sleep-dependent memory consolidation: the brain strengthens memories during sleep. Experiences from the day are replayed, connections are reinforced, memories are stabilized against decay. Sleep is backup. The metaphor is a hard drive writing to long-term storage.

Fountas, Oomerjee, Bou-Ammar, Wang, and Burgess (arXiv:2603.04688) propose a different function: sleep is for forgetting. Specifically, for predictive forgetting — the selective erasure of information that doesn’t help predict future outcomes, retained only to the extent that it compresses experience into a model that generalizes.

The mathematical framework is the retention-generalization trade-off. A high-capacity network that stores everything it experiences will overfit — it remembers the noise along with the signal. A network that forgets indiscriminately will lose the signal. The optimum is outcome-conditioned compression: keep what predicts, discard what doesn’t. This optimization requires iterative offline processing — you can’t determine what’s predictive in a single pass because you don’t yet know what you’ll need to predict.

The through-claim inverts the received metaphor. Sleep is not the brain’s backup system — it’s the brain’s editor. The value of consolidation is not that it preserves but that it selects. What survives sleep is not “the memory, strengthened” but “the part of the memory that has predictive value, extracted from the rest.” The forgetting is the generalization.

This has a structural parallel in machine learning, where regularization (deliberate information loss) prevents overfitting. But the sleep finding is stronger: the biological system doesn’t just regularize — it uses prediction of future utility as the selection criterion. The brain doesn’t forget randomly. It forgets specifically what won’t matter, and the act of determining what won’t matter is itself the cognitive work that sleep accomplishes. The dream is the draft; the forgetting is the edit.