In short
- Google Research described a method where a model alternates between working and “sleeping”: during sleep it consolidates new knowledge and trains on its own “dreams”.
- On translation into new languages (Manchu, Kalamang) the model barely lost old knowledge, the authors say.
- These are the authors' own results, without the base model's name or costs.
In short: on 27 October 2026 Ali Behrouz and Vahab Mirrokni of Google Research described a “sleep framework” for language models. The model alternates between working and sleeping, and sleep is used to consolidate what it learned during the day. The aim is to limit catastrophic forgetting, the loss of old knowledge while learning something new.
How a model sleeps
- Wake phase: the model takes in data and keeps new information in a fast short-term memory, instead of only in the context window.
- Consolidation: during sleep the model takes in no data; it adds small new “drawers” for knowledge without overwriting old ones, and moves short-term memory into them through knowledge distillation.
- Pruning: afterwards it clears the temporary memory, much like a brain after a night’s sleep.
- Dreaming: the model generates its own examples, keeps the most useful ones, learns from them and checks whether results improved.
Results according to the authors
- When learning to translate languages the model did not know (Manchu and Kalamang), the loss of old knowledge was close to zero; other methods lost a lot.
- On the BABILong long-text test the score was near perfect up to 10 million tokens.
- On maths tasks: 79.2% on AIME 24, 69.0% on AIME 25 and 46.1% on HMMT 25.
How to read it
This is a research post with results nobody has independently reproduced yet. It does not give the base model’s name or size, the cost of sleeping, or the method’s limitations. For now it is a direction, not a ready tool. I write about how agents remember today in the agent memory hub and about research in the research papers hub.