Deployed Models Never Learn. Google’s ‘Sleep’ Framework Gives Them a Memory Cycle

A deployed LLM is frozen. It has a knowledge cutoff, it answers from a context window, and when the session ends everything it learned is gone. Post-training and retrieval both look like answers to that, and neither is: one has to be scheduled, the other just moves the problem.

Google Research’s latest write-up describes a third option — treat the model as having a sleep cycle.

Wake, Then Sleep

Under the sleep paradigm, the model alternates between an active phase, where it handles inputs and holds temporary information in fast short-term memory, and an offline sleep phase, where it stops serving traffic and turns its compute inward.

The architecture makes this possible: a Mixture-of-Experts model with pre-allocated memory blocks. During sleep, newly unlocked low-rank experts are integrated into stable, lower-frequency memory modules, so new knowledge lands in expanded capacity rather than overwriting old parameters. That is the direct answer to catastrophic forgetting — not better regularization, more room.

Fragile in-context memories are then moved down into those parameters through knowledge seeding, an upward distillation combining Generalized Knowledge Distillation with RL-based imitation learning. Once consolidation finishes, the temporary high-frequency parameters are reset — synaptic pruning, freeing capacity for the next session.

Dreaming

The second stage is self-improvement. The MoE router randomly activates non-standard experts so the model mixes domains it would not otherwise combine, generates synthetic training samples, filters them by gradient-based data selection, and fine-tunes on the survivors. The theory section frames this as broadening the set of gradient directions explored, not as ordinary data augmentation.

What It Actually Bought

On BABILong, sleep-based consolidation scales to 10M tokens, past GPT-4 and Google’s Titans. On factual incorporation it reached 48.9% single-task and 46.2% on continued pre-training, beating SFT and SEAL. On AIME 24/25 and HMMT 25, sleep-enhanced base models hit 79.2%, 69.0% and 46.1% over SFT, GRPO and on-policy self-distillation. In continual translation into unseen languages such as Manchu and Kalamang, the framework held a near-zero performance drop where ICL, Cartridges and SFT collapsed.

The caveat is the one the paper states plainly: “sleep” runs inside a fixed architecture and human-defined training rules. Nothing here is self-directed in any meaningful sense.