Google's sleep framework lets LLMs continually learn and retain up to 10M tokens
iruletheworldmo · x · 2026-11-01
Google Research's Ali Behrouz and Vahab Mirrokni propose a sleep framework for continual learning, inspired by how the human brain consolidates memory during rest. LLMs alternate between wake phases (active interaction) and sleep phases (offline memory consolidation), turning fragile short-term context into stable long-term parameters and moving beyond static knowledge cutoffs.\n\nThe framework tackles two pain points: continual parameter updates cause catastrophic forgetting, while pure in-context learning loses knowledge once the session ends. Experiments show the approach lets models continuously acquire new facts, retain long-context knowledge up to 10M tokens, and improve complex reasoning.
Related event: Google's Sleep Framework Enables Continual Learning for LLMs(2 posts)→
More from Models
- DeepSeek v4.1-flash not secretly updated: API fingerprint unchanged since Oct 27 — teortaxesTex · 2026-11-01
- OpenAI caps paid ChatGPT accounts per device and payment card — The Decoder · 2026-11-01
- 'Claudese' is back: side-by-side screenshot shows Claude's signature prose style returning — oyacaro · 2026-11-01
- Same model, 17 points apart: OpenCode hits 82% on 21 MCP servers vs Claude Code's 65% — alejxb · 2026-11-01
- Commenter: if GPT-6.2 can't beat Opus 5.5, OpenAI faces a long road back — CtrlAltDwayne · 2026-11-01
- DeepSeek quietly drops autonomous coding agent prototype, Bloomberg picks it up instantly — teortaxesTex · 2026-11-01