Google's 'Sleep Framework' helps AI models keep learning after launch

Most big AI models stop learning once training ends, so their knowledge has a cutoff date. New info from chats gets lost, and retraining can wipe out old skills. Google's research team unveiled a Sleep Framework for continual learning, inspired by how human brains lock in memories during sleep. The model runs in wake and sleep phases: it takes in new info while awake, then sorts and improves its memory while asleep. In the sleep phase, it moves short-term info into long-term memory using parameter expansion, expert memory blocks, and knowledge distillation. It also dreams by making its own practice samples to sharpen reasoning. Tests show it handles up to 10 million tokens of context, learns new languages with almost no loss of old skills, and scores higher on math tests like AIME 24/25.