Back to Project
G

Google sleep framework

Project Oct 29, 2026

A research framework that alternates active interaction with offline consolidation and synthetic rehearsal to retain knowledge in model parameters.

Visit primary link 1 source attached
Google Research diagram showing an active input phase, short-term memory, offline consolidation and dreaming, and consolidated parametric memory.
Google Research's proposed sleep-phase architecture, reproduced from its October 27 explanation. This is a research diagram, not a deployed Gemini feature or an AITrending reproduction of the experiments. Source: Google Research
ai-agentsverification

What happened lately

Google sleep framework timeline

Google Research’s sleep framework separates an active phase, which accumulates short-term information, from an offline phase that consolidates selected knowledge into model weights. Its October 27, 2026 explanation describes pre-allocated mixture-of-experts memory capacity, distillation and reinforcement learning, followed by synthetic rehearsal called dreaming.

The authors report evaluations of continual translation, factual knowledge, mathematical reasoning and long-context retention. These are research results attributed to the team. The blog does not establish that a consumer Gemini product automatically remembers every interaction or offers the experimental long-context result as a live context window.