LLMs Consolidate Memory Through "Sleep"
behrouz_ali · x · 2026-07-14
A paper titled "Language Models Need Sleep: Learning to Self-Modify and Consolidate Memories" suggests that while LLMs can learn within their context window, they lose this knowledge once the session ends.
The paper introduces a "sleep" phase:
- First, consolidate short-term memory into long-term weights
- Then, "dream/review" by generating RL-trained synthetic data
The authors describe this mechanism as a path to continuous learning: models are no longer completely frozen after pre-training. Instead, they can self-modify, reduce catastrophic forgetting, and perform better on continuous learning and reasoning tasks.
More from Research
- arXiv moderator's tip: accept the classifier's primary category or risk a hold — tdietterich · 2026-10-27
- arXiv has 4,000+ papers on hold amid submission surge; moderator shares how to avoid it — tdietterich · 2026-10-27
- CoreWeave, You.com and NVIDIA lift BrowseComp 22.9% in 3-week RL post-training sprint — RichardSocher · 2026-10-27
- Open-source decision models sold their own base in Red Alert 2 — then learned until no human could beat them — heathercmiller · 2026-10-27
- DeepMind paper: LLMs can't sample randomly on their own — pass random numbers from code instead — rohanpaul_ai · 2026-10-27
- World Model Assigns Weight and Friction to Objects, Then Simulates with a Physics Engine — jonstephens85 · 2026-10-27