Researchers have a new pitch for making chatbots remember you without a bigger context window: stop writing what you tell it into the prompt, and start writing it into the model's weights.
The idea, described in an arXiv paper posted September 17, is called the Infinite-Parameter LLM. A small "hypernetwork" takes whatever data shows up during a session, like facts you provide or corrections you make, and turns it into a low-rank tweak to a shared base network's feed-forward weights. Instead of reading that context once and locking it in, the system keeps a running probabilistic estimate of what it should know and updates the weights as the conversation continues. The stored model stays the same size the whole time. What changes is which version of the weights gets compiled for you.
This matters because retrieval-augmented generation and long-context prompting are both patches for the same underlying problem: today's models can't actually learn during a conversation, they can only re-read. Every fact you feed a chatbot gets discarded the moment the session ends, and every turn pays the compute cost of re-processing it. If weight generation works as described, it trades that per-turn re-reading tax for a one-time update, frees up context window space for other things, and could let personalization survive between sessions instead of vanishing.
The catch is that this is a single paper with a proposed evaluation protocol, not a benchmarked result. Mixture-of-Experts architectures took years and enormous compute budgets to prove out at scale; a hypernetwork that reconstructs weights from a Bayesian belief state is a more exotic bet, and the paper does not yet report how it performs against the in-context and retrieval baselines it aims to beat.