AdaMem: Learning What to Remember for Personalized Long-Horizon LLM Agents

2026-06-19

AI Summary

This work addresses the challenge of memory bloat in long-term memory systems for personalized large language model (LLM) agents, which often arises from indiscriminate recording and degrades question-answering accuracy. The authors propose AdaMem, a novel approach that formulates memory write control as a learnable, role-dependent policy and dynamically refines memory content using weekly-granularity user feedback. Key innovations include a structured memory policy, a lightweight patch-based self-reflection mechanism, a failure rollback protocol, and a feedback-driven policy optimization framework, accompanied by the introduction of AdaMem-Bench, a dedicated evaluation benchmark. Experimental results demonstrate that AdaMem improves question-answering accuracy by up to 9.0% over the Mem0 baseline while simultaneously reducing memory footprint by 9%, across two retrieval models and feedback modalities.

Abstract

Long-term memory systems for Large Language Model (LLM) agents typically try to \emph{remember everything}, extracting memories uniformly to retain as many facts as possible. In production, however, inference cost and finite context budgets make this untenable: beyond consolidating raw dialogue into memory, an agent must exert \emph{write control}, efficiently keeping only the information each user actually cares about. Otherwise, long-horizon personalized interactions suffer \emph{memory bloat}, where irrelevant trivia crowds out useful information and steadily erodes question-answering (QA) accuracy. We argue that what is worth remembering is role-dependent, and propose \textbf{AdaMem} (Adaptive Memory), a method that \emph{learns what to remember} for each user from feedback. AdaMem maintains a structured, role-specific Memory Policy and refines it from weekly QA feedback through a lightweight, patch-style self-reflection step with failure rollback. To study this setting, we build \textbf{AdaMem-Bench}, a benchmark that simulates weeks of interaction with week-by-week QA. Across two extraction models and two feedback modes, AdaMem improves QA accuracy by up to \textbf{+9.0\%} over the uniform Mem0 baseline while shrinking memory volume by \textbf{9\%}.

Problem

• long-horizon LLM agents
• personalized memory
• memory bloat
• write control
• question-answering accuracy

Innovation

• adaptive memory
• memory policy
• personalized LLM agents
• feedback-driven learning
• memory bloat mitigation