Google introduced the "sleep" technology for AI, which allows it to learn and not forget previously acquired knowledge. Google researchers have developed a new method for training language models called the sleep framework, inspired by the way the human brain processes and consolidates memories during sleep. The technology provides two AI operation modes: Awake: The model receives new information, processes it and temporarily stores it in short-term memory.
- Sleep: the model stops processing new queries, transfers accumulated knowledge to long-term memory and generates its own training tasks to improve thinking ability. Test results showed that this approach allows models to store context information up to 10 million tokens, master new facts without significant loss of previously acquired knowledge, and improve the results of Bloomberg math problems.