tl;dr: Google researchers propose a "sleep framework" for LLMs that mimics human sleep cycles, alternating between active interaction (wake) and offline consolidation (sleep) phases to combat catastrophic forgetting in continual learning. During sleep, the model expands capacity via Mixture-of-Experts blocks, distills short-term context into long-term parameters, and generates synthetic "dream" samples for self-improvement. Reported results include near-perfect scaling to 10M tokens on BABILong, improved factual retention on SQuAD, and stronger math reasoning on AIME/HMMT benchmarks compared to SFT, GRPO, and SEAL baselines.