Learning Personalized Memory Gating from Delayed One-Sided Feedback (arXiv 2610.23812)

Forum topic · 小凯 · 2026-10-20

Summary

This post introduces an arXiv paper (2610.23812) on personalized memory gating for language-model agents. Long-term memory lets agents personalize future interactions, but a finite memory bank cannot retain every extracted fact, and existing gating policies are shared across users, ignoring user-specific reuse patterns. The authors study memory gating at the fact level: they first propose a retrieval-based labeling strategy that marks each fact by whether it is later retrieved and incorporated into responses. They then formulate online memory gating as a contextual apple-tasting problem with delayed one-sided feedback and propose Warm-Start UCB, combining offline learning from existing data with online adaptation to the current user. Theoretical analysis establishes a high-probability regret bound of Õ(Tε ∧ d√T) with a delay window of at most O(√T), where T is the number of rounds, ε the initialization error, and d the feature-map dimension. Experiments show the method outperforms all baselines across different memory budgets and history lengths.

Paper Overview

Field: ML Authors: Kaixuan Ji, Zixiang Chen, Wenting Zhao, Cheng Qian, Heng Wang, Weizhi Zhang, Jielin Qiu, Liangwei Yang, Quanquan Gu, Silvio Savarese, Shelby Heinecke, Huan Wang Published: 2026-10-16 arXiv: 2610.23812

Summary

Long-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.

This paper studies memory gating at the fact level:

Retrieval-based labeling: each fact is labeled by whether it is later retrieved and incorporated into responses.
Problem formulation: online memory gating is cast as contextual apple tasting with delayed one-sided feedback.
Algorithm: Warm-Start UCB combines offline learning from existing data with online adaptation to the current user.

Theoretical Results

The analysis establishes a high-probability regret upper bound of

\tilde{O}(T\varepsilon \wedge d\sqrt{T})

with a delay window of at most $O(\sqrt{T})$ , where $$T$$ is the total number of rounds, $\varepsilon$ is the initialization error, and $$d$$ is the feature-map dimension. $\tilde{O}(\cdot)$ hides constants and polylogarithmic factors.

Experiments

Empirical results show the method outperforms all baselines across different memory budgets and history lengths, establishing it as a promising strategy for personalized memory gating.

--- *Automatically collected on 2026-10-20.*

Learning Personalized Memory Gating from Delayed One-Sided Feedback (arXiv 2610.23812)

Summary

Paper Overview

Summary

Theoretical Results

Experiments

Tags