Google's Sleep Framework Lets AI Models Keep Learning Without Forgetting
Google Research proposes a biologically inspired sleep phase that lets LLMs consolidate new facts into weights without forgetting, scaling context to 10M tokens.
- Google Research introduced a sleep framework letting LLMs consolidate knowledge offline between wake sessions.
- Alternates active inference with offline memory consolidation and dreaming-based self-improvement cycles.
- Uses MoE parameter expansion plus on-policy distillation to avoid catastrophic forgetting during updates.
- Scales to 10M-token long-context tasks on BABILong, beating GPT-4 and Titans baselines.
- Hits 79.2% AIME 24, 69.0% AIME 25, 46.1% HMMT 25 on math reasoning.
- Full method detailed in the arXiv paper from Behrouz, Hashemi, Javanmard and Mirrokni.
Google Research has proposed a continual learning framework that lets a deployed large language model alternate between serving requests and consolidating new information. The neuroscience-inspired Google Research overview describes a “wake” phase for processing inputs and a “sleep” phase for updating long-term memory, practicing on synthetic data, and preparing temporary capacity for reuse.
Deployed LLMs usually retain the knowledge encoded during pre-training and subsequent fine-tuning. Context windows and retrieval systems can supply newer information at inference time, but those additions disappear when the prompt or external source goes away. Repeated fine-tuning can also cause catastrophic forgetting, in which learning a new task degrades performance on earlier ones. Google’s framework addresses both limits by scheduling weight updates throughout the model’s operating life.
Two phases, one model
During the wake phase, the model processes inputs and stores recent information in fast, temporary memory components. The sleep phase pauses new input processing and uses that accumulated material for memory consolidation and self-improvement.
The framework keeps the model’s overall architecture fixed while unlocking capacity reserved in advance. It uses a Mixture-of-Experts architecture, which routes computation through selected parameter blocks called experts. Newly activated, low-rank experts provide compact areas for storing incoming knowledge. This design makes sleep an automated post-training cycle within deployment.
From temporary traces to weights
The sleep stage contains two processes inspired by non-rapid eye movement and rapid eye movement sleep. The NREM-inspired process consolidates short-term information through three mechanisms:
- Parameter expansion: The model unlocks pre-allocated, low-rank experts and integrates them with stable, lower-frequency memory blocks. Fresh capacity reduces interference with older parameters.
- Knowledge seeding: Generalized Knowledge Distillation transfers information from short-term context into the new experts. On-policy distillation trains the expanded model on sequences it generates itself, using token-level distributions from a teacher model. A Learning-to-Imitate reward compares generated and teacher completions through semantic similarity and Levenshtein edit distance.
- Synaptic pruning: After consolidation, the framework resets high-frequency temporary parameters so they can store information from later wake phases.
Knowledge seeding treats memory consolidation as low-rank compression. The model identifies a compact set of parameter directions that preserve useful behavior and semantic information, then writes those directions into newly available capacity. The approach aims to reduce the parameter interference that drives catastrophic forgetting.
Dreams become training data
The REM-inspired process generates synthetic training examples. The Mixture-of-Experts router activates unusual combinations of experts, causing the model to combine knowledge from domains that rarely interact during normal inference. Those combinations produce candidate prompts, solutions, or reasoning traces for self-training.
A gradient-based selection method scores the candidates according to their expected training value. The model trains on the highest-ranked samples and discards weaker ones. Google’s researchers argue that varied expert routing broadens the gradient directions explored during optimization, while filtering concentrates compute on examples with a stronger learning signal.
Reported results
Google evaluated the framework on continual translation, long-context recall, factual integration, and mathematical reasoning. The authors report lower forgetting during sequential language learning and substantial gains over the baselines included in the study.
| Evaluation | Reported result | Comparison |
|---|---|---|
| Continual translation | Near-zero performance loss across multi-stage evaluations in unseen languages including Manchu and Kalamang | Lower forgetting than in-context learning, Cartridges, and supervised fine-tuning |
| BABILong | Near-perfect performance through sequences representing up to 10 million tokens | Higher reported performance than GPT-4 and Titans baselines |
| Factual integration | 48.9% on single-task evaluation and 46.2% after continued pre-training | Higher than the base model, supervised fine-tuning, and Self-Adapting Language Models |
| AIME 2024 | 79.2% | Higher reported scores than supervised fine-tuning, GRPO, and on-policy self-distillation |
| AIME 2025 | 69.0% | |
| HMMT 2025 | 46.1% |
The BABILong result measures consolidation across extremely long inputs. The model processes information during wake periods and compresses it into parametric memory during sleep; it does not attend to all 10 million tokens in one conventional inference window. That distinction affects comparisons with models whose advertised context length refers to tokens available within a single forward pass.
The reported experiments establish performance within the paper’s benchmark settings. They leave open questions about independent replication, training stability over many sleep cycles, sensitivity to reward quality, and behavior on continuously changing production data.
What deployment would require
A production implementation would need an operating schedule and infrastructure for both phases. Several constraints follow from the design:
- Reserved model capacity: The method assumes a Mixture-of-Experts backbone with expert slots allocated before deployment. Dense models would require architectural changes or another mechanism for adding isolated capacity.
- Offline compute: Consolidation, synthetic generation, gradient scoring, and weight updates consume accelerator time during sleep cycles. The work moves outside the request path, though serving latency still depends on the underlying MoE architecture and deployment topology.
- A reliable evaluator: Dream filtering requires a reward model or downstream scoring function. Weak evaluation can select low-quality synthetic examples and reinforce model errors.
- Versioning and rollback: Every sleep cycle changes model weights. Production systems would need checkpoints, regression tests, staged releases, and a way to restore an earlier state.
- Data controls: Consolidating user interactions into weights creates privacy, retention, and poisoning risks. Implementations would need strict input filtering and policies governing which information may enter long-term memory.
- Fleet coordination: Systems serving traffic across many replicas would need to synchronize updated experts or rotate replicas through sleep without fragmenting model behavior.
Retrieval-augmented generation would remain useful for source attribution, access control, frequently changing records, and information that should stay outside model weights. Sleep-style consolidation targets durable knowledge and capabilities that justify the cost and governance burden of a parameter update.
A research path, not a service
Google presents the framework as research and has announced no production offering. The Sleep paper describes the full method. Its memory system builds on the Hope architecture paper and Google’s earlier Titans overview.
The design turns model adaptation into a scheduled operational process. If the results transfer to production workloads, teams could manage wake and sleep cycles alongside evaluation, deployment, monitoring, and rollback, giving continually updated weights the same controls applied to other changing production artifacts.
Comments
