Google proposes 'sleep framework' for LLM continual learning

The model alternates between active input processing and sleep phases, where memory is stabilized via MoE-based expansion and self-generated synthetic data. Parameter scaling, on-policy distillation, and RL refinement allow continuous adaptation without catastrophic forgetting, supporting up to 10M token context retention.

Summarised from the original article by an open-source model, then checked automatically: every number and English name must appear in the source. The original article is authoritative.