Google Research’s Sleep framework learns new knowledge temporarily during a waking phase and consolidates it during a sleeping phase. It progressively activates preallocated experts, using distillation, reinforcement learning and synthetic rehearsal to reduce loss of earlier capabilities. Researchers report improved retention in long-context evaluations and sequential translation tasks. These results come from a fixed architecture and designed procedure; they do not mean an AI creates its own code or unlimited memory. The work addresses evaluation of continual learning rather than announcing a generally available commercial feature.