Warm-Start UCB personalizes memory gating for language agents
Original title: Learning Personalized Memory Gating from Delayed One-Sided Feedback
The policy learns which facts to keep from delayed retrieval feedback and beat baselines across memory budgets.
The paper studies fact-level memory gating for language-model agents that cannot store every extracted fact. It labels facts by whether they are later retrieved and used, then frames online gating as contextual apple tasting with delayed one-sided feedback. Warm-Start UCB combines offline learning with online adaptation to the current user. The authors give a high-probability regret bound when the delay window is at most O(sqrt(T)) and report that the method beat all baselines across memory budgets and history lengths.
Unconfirmed: Baseline gains are reported only by the authors' experiments.
- Existing shared gating policies ignore user-specific reuse patterns.
- Labels come from later retrieval and use in responses.
- The stated regret bound is tilde-O of T times epsilon and d times square root of T.
From the original article
1 min read at arxiv.orgLong-term memory enables language-model agents to personalize future interactions, but a finite memory bank cannot preserve every extracted fact. Existing memory gating policies are typically shared across users and therefore fail to account for user-specific patterns of future reuse. In this paper, we study memory gating at the fact level.
By Kaixuan Ji · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.