Google Research's sleep framework transfers short-term context into longer-lasting knowledge in model weights. Google published its explanation on October 27, 2026. The approach addresses catastrophic forgetting: learning new information can degrade performance on earlier tasks. The results discussed here are the researchers' reports, not independent measurements.
The publication history matters. The paper's arXiv v2 text says a version was public on OpenReview in September 2025. Google's October post therefore explains research with an earlier public version; it does not establish that the method first appeared that month.
What happens during sleep
During active use, the model receives information and stores it in temporary memory. In the offline phase, it pauses new inputs, activates pre-allocated mixture-of-experts memory blocks and updates its numerical weights using self-generated practice examples, according to Google's explanation. The architecture and training rules remain human-defined.
The paper separates the process into consolidation and dreaming. Consolidation transfers knowledge into additional parameter capacity. Dreaming generates synthetic training material to rehearse information and refine capabilities. For developers evaluating persistent-memory claims, the distinction is practical: this method changes model parameters. Keeping conversation notes available for the next session would not, by itself, demonstrate that mechanism.
Read the 10-million-token result with its limitation
Google reports near-perfect scaling to 10 million tokens on BABILong, a long-context benchmark. Its explanation attributes that result to consolidating long-context information into parameters. The number describes a research benchmark result, rather than an announced context-window size for a commercial chatbot.
The paper's appendix B.2 supplies an important limitation: all small models, including Hope, suffered substantial performance drops without fine-tuning. That finding limits what the headline result establishes about performance without adaptation.
When assessing follow-up claims, check whether the model received fine-tuning and what performance the researchers report without it. The supplied evidence supports the degradation finding; it does not establish the exact fine-tuning protocol used for the BABILong result.