Language-model agents can use long-term memory to personalize future interactions, but a finite memory bank cannot preserve every extracted fact. The challenge is deciding which facts to keep and which to discard—a process known as memory gating.
Existing memory gating policies are typically shared across all users, which the authors argue is a key limitation. Because different users have different needs, a one-size-fits-all policy may not retain the most relevant information for each individual.
The paper introduces an approach to learn personalized memory gating policies from delayed one-sided feedback. This suggests a method that adapts gating decisions per user, using feedback that may arrive after a delay and only from one direction (such as positive or negative signals). The details of the method are not fully described in the abstract, but the direction is clear: moving from shared to personalized memory management.