Researchers propose weight-based memory system that lets LLM user data fade
An arXiv paper describes a fixed-capacity adapter plus a learned controller that stores, updates and expires user memories inside model weights, reporting 86% of oracle performance on a purpose-built evaluation dataset.
In plain words
Large language models usually remember user information either in their weights or by retrieving it from a database. The paper's authors built a system where a small adapter inside the model stores facts, and a controller decides what to keep, update or let expire. In tests, it recalled valid facts and stayed silent about expired ones better than existing approaches.
What was announced
The authors argue that existing parametric memory approaches inherit the goal of continual learning, which is to retain information indefinitely, and that this fits poorly with memory, which must keep stable facts, reinforce recurring observations, update changed beliefs and let transient information fade. Their system pairs a memory substrate with a learned policy: the substrate is a fixed-capacity LoRA adapter using k-hot rank combinations to store facts without catastrophic interference, with an independently controlled temporal decay so outdated information fades while long-term facts persist. A lightweight controller governs each incoming user event so the memory reflects the user's updated belief state.
For evaluation, the team built a controlled dataset of long-horizon user trajectories in which facts vary in persistence as beliefs and behavior shift, paired with probe queries testing recall while information is valid and abstention once it expires. They report that the trained policy reaches 86% of oracle performance by balancing recall and abstention, while continual learning and retrieval-augmented generation baselines mostly default to one behavior or the other. The paper positions itself as the first system to manage memory directly in model weights.
What it does not show
- The evidence is limited to the paper's abstract, so the method's details, experimental setup and reported results cannot be independently verified from the supplied passages.
- The 86% of oracle performance figure is the authors' own reported result on their own dataset and has not been confirmed by outside evaluation.
- No information is given about deployment, model sizes, computational cost or real-world user data.
Sources
Source-support review completed 21 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 21 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.