When an AI learns something useful in one chat, why doesn't every other user's chat get smarter too?
What prevents individual session learning from scaling system-wide?
This explores why something a model picks up during one conversation or agent run (in-context learning, per-user adaptation) doesn't automatically become knowledge the whole system shares.
This explores why something a model picks up during one conversation or agent run doesn't automatically become knowledge the whole system shares. The corpus has no note on pooling what many users' sessions teach a model, so what follows is assembled from neighbouring work. It points to three obstacles: consolidation is expensive, writing into shared weights risks erasing things, and much session-level learning is built to stay local.
The first obstacle is compute. What a model learns mid-session sits in its context, which is temporary. One note argues the long-context bottleneck isn't memory capacity. It's the compute needed to turn evicted context into internal state, and quality keeps improving with more consolidation passes (Is long-context bottleneck really about memory or compute?). A companion proposal is a "sleep phase". The model distills what it learned in context into its weights offline, using RL-generated "dreams" as rehearsal (Can models consolidate memories during offline sleep phases?). Session learning doesn't scale partly because turning it into durable knowledge is a separate, costly job that doesn't happen while the session is running.
The second obstacle is forgetting. Writing new material into weights can overwrite old material, which is why the sleep paradigm is framed as consolidating "without forgetting". Two notes sidestep the problem by never touching weights. VOYAGER stores executable skills in an embedding-indexed library and composes new skills from old ones (Can agents learn new skills without forgetting old ones?). AgentFly reaches 87.88% on GAIA validation using case, subtask and tool memories with no parameter updates (Can agents learn continuously from experience without updating weights?). Both are real continual learning, but the knowledge lives in that agent's external store rather than in the model. Sharing it means sharing the store, which is a different problem from improving the model itself.
The third obstacle is that a lot of session-level learning is designed to stay local. PReF personalizes a user from about ten adaptive questions at inference time, again without changing weights (Can user preferences be learned from just ten questions?). What it learns is a small set of coefficients for that person, not a lesson for the model. Titans-style neural memory keeps whatever is surprising to the current sequence (Can neural memory modules scale language models beyond attention limits?). Surprising to one session isn't necessarily worth keeping for everyone, though that step is my inference rather than something the note tests. Session signal is also often enough by itself. A single-session encoder rivals recommenders that use richer cross-session history (Can single sessions alone rival history-rich recommendation?), so nothing forces the learning to travel.
There is one more possible catch. Moving knowledge between networks is itself fragile. Distillation becomes unstable when teacher and student are far apart, and a nearby "proximal" teacher with a trust region fixes that (Can proximity between teacher and student fix distillation instability?). The sleep approach relies on distillation, so a session-adapted model drifting away from the shared model may hit the same instability. That link is a hypothesis, not a finding. What the corpus doesn't cover is the social side of scaling: whose sessions to learn from, how to resolve conflicting lessons, and how to protect privacy.
Sources 8 notes
Research shows the bottleneck is not memory capacity but the compute required to consolidate evicted context into fast weights during offline sleep phases. Performance improves with more consolidation passes, following a test-time scaling pattern on harder reasoning tasks.
The Sleep paradigm uses Knowledge Seeding (distilling smaller networks into larger ones) and Dreaming (RL-generated rehearsal) to consolidate in-context knowledge into weights without forgetting. Gains appear in long-context understanding, few-shot reasoning, and continual learning.
VOYAGER demonstrates that storing executable skills in an embedding-indexed library and composing complex skills from simpler ones allows agents to learn continuously while avoiding the forgetting that occurs with weight-update-based methods. Environmental feedback refines skills while an automatic curriculum drives continual exploration.
AgentFly formalizes agent learning as a Memory-augmented MDP with three memory modules (case, subtask, tool) that enable credit assignment and policy improvement entirely through memory operations. The approach achieved 87.88% on GAIA validation without modifying LLM parameters.
PReF learns base reward functions from preference data, then uses active learning to select maximally informative questions that reduce coefficient uncertainty. Users can be personalized via inference-time reward alignment without weight modification.
Show all 8 sources
Titans architecture separates attention (short-term, quadratic) from neural memory (long-term, compressed), prioritizing surprising tokens for storage. The model outperforms standard Transformers and linear RNNs across tasks while scaling to 2M+ token contexts without quadratic penalties.
Sequential Masked Modeling adapts encoder-only transformers for session-based recommendation using penultimate-token masking and sliding-window augmentation. Across three datasets, this single-session approach consistently outperforms other single-session methods and rivals cross-session approaches with richer user history.
TOP-D constructs a close teacher instead of distilling from a distant target, bounded by a trust region. This controls gradient variance, guarantees monotonic improvement, and outperforms standard distillation with zero computational overhead.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Language Models Need Sleep
- SkillRL: Evolving Agents via Recursive Skill-Augmented Reinforcement Learning
- Continual Learning Bench: Evaluating Frontier AI Systems in Real-World Stateful Environments
- Native Sparse Attention: Hardware-Aligned and Natively Trainable Sparse Attention
- Useful Memories Become Faulty When Continuously Updated by LLMs
- AgentFly: Fine-tuning LLM Agents without Fine-tuning LLMs
- Memorization and Knowledge Injection in Gated LLMs
- Optimizing Encoder-Only Transformers for Session-Based Recommendation Systems