Google tests a sleep phase that lets a language model keep new facts

Google Research published a sleep method for large language models (LLMs, AI that writes text) on 27 October 2026. A model in use stays fixed, and it forgets new facts when a session ends. In the sleep phase, the model updates its weights, the numbers that hold what it knows, with practice examples that it makes itself.

How it works

The model has two phases: active use and offline sleep. In sleep, the model opens memory blocks that were set aside before and trains on its own practice examples. The model architecture stays the same.

The numbers

The authors report these results:

• 48.9% on a test of new facts, above other methods
• 79.2% on the AIME 24 maths test
• near-perfect results with 10M tokens of long context

What is not known

All results come from the authors. The source gives no cost, no time for a sleep phase and no plan to sell it. It gives no test by an external group.

Insight: Where a model keeps what it learns

llm-d 1.0, open-source software that operates models on many servers, keeps the work of a model in a cache. Then the model does not do the same work again. The Google paper puts new knowledge into the weights. This reads as two answers to one question: does a model keep what it learns in a cache or in its weights? Builders must compare the cost of each.