Personalize-then-Store: Benchmarking and Learning Personalized Memory for Long-horizon Agents

Yeonjun In, Wonjoong Kim, Sangwu Park, Kanghoon Yoon, Chanyoung Park

arXiv:2605.25535 · 2026-05-27 공개 · arXiv · PDF

long-horizon-agents llm-memory personalized-memory permembench memory-personalization storage-gating session-gating

Abstract

Existing large language model (LLM) based memory systems apply universal, static policies that overlook a fundamental reality: the contexts that are worth storing in memory are different across users. This misalignment wastes limited memory budget on transient interactions while failing to preserve critical context for long horizon tasks. To address this gap, we investigate an underexplored question: can LLM based memory systems learn personalized memory policies? We introduce PerMemBench, the first benchmark for evaluating personalized memory systems, featuring multi year, multi domain interaction histories across diverse user personas. We further present the first empirical study of memory personalization, proposing session level storage gating, a lightweight framework that selectively bypasses memory operations for transient sessions. Our study confirms that personalization yields substantial retention gains under perfect gating, yet reveals that accurate gating remains an open and critical challenge.

한국어 요약

📋 한 줄 요약

**[LLM Memory / Personalization]** PerMemBench이 multi-year·multi-domain·diverse persona의 첫 personalized memory 벤치마크 — session-level storage gating으로 transient session bypass 시 perfect gating 하에 substantial retention gain, 정확한 gating은 open challenge.

🎯 핵심 기여도

💡 핵심 아이디어

LLM memory 시스템의 한정 budget을 사용자별로 다른 "저장 가치 있는 context"에 맞춰 분배해야 하며, session-level storage gating으로 transient session을 bypass하는 personalization은 perfect gating 하에 substantial retention gain을 제공하지만 정확한 gating 자체가 critical open challenge다.

🔬 기술적 접근법

📊 주요 결과

💭 의의 및 한계

**의의**: Memory 시스템의 universal-static 가정 반박, personalization이라는 새 차원 도입, perfect gating의 substantial gain으로 잠재력 정량화, 첫 벤치마크로 분야의 evaluation 인프라 제공. **한계**: 정확한 gating 알고리즘 미해결, "perfect gating" 가정의 실용 적용 격차, multi-year·multi-domain 데이터 큐레이션의 일반화, abstract에 구체 수치 미포함.

🚀 실용적 활용