What is SLEEP?
SLEEP is a Google Research framework for continual learning in large language models. It aims to reduce the loss of previous capabilities as new knowledge is acquired.
Why explore a new approach?
Updating a model with new data can degrade previously learned skills. Relying only on the conversation context also limits retention to the available context window.
What happens in the sleep phase?
The framework combines active interaction with a knowledge-consolidation phase. Sleep and dreaming are metaphors for technical processes; they do not imply human sleep or consciousness.
What do the results show?
Google reports comparative results on continual-learning and long-context evaluations. The charts apply to specific datasets and experimental conditions; they do not guarantee the same performance in every model.
When can it be used?
This is a research approach, not evidence that every chatbot now has permanent memory or a ten-million-token real-world conversation capacity. Deployment and practical performance require separate evaluation.
