Google's "sleep framework" lets LLMs learn continuously with zero forgetting

HowToPrompt__ · x · 2026-11-18

Google Research published a paper introducing a biological "sleep framework" for LLMs to overcome catastrophic forgetting. The model alternates between a Wake Phase, where new experiences land in fragile short-term memory (KV cache), and a Sleep Phase, where offline consolidation uses low-rank structures and a Mix of Experts to transfer short-term tokens into permanent parametric memory. The sleep cycle also includes "dreaming": the model deliberately activates random expert routing, generates synthetic scenarios to self-test reasoning, and refines them with reinforcement learning. Reported results: perfect retention across a 10-million-token window, learning unseen languages with zero performance drop, and 80% accuracy on AIME. The takeaway: the future isn't bigger context windows, it's letting machines process experiences after you log off.

Related event: Google's "Sleep Framework" Paper Sparks Debate(2 posts)→

Original post →

More from Models

Models channel →