The sleep framework: A new way for LLMs to retain knowledge in continual learning

The sleep framework equips LLMs with a periodic lifecycle of active interaction and offline memory consolidation, enabling continual learning and long-context retention up to 10M tokens.