Dynamic Affective Memory Management for Personalized LLM Agents

TL;DR

This work tackles the challenge of long-term affective memory in personalized LLM agents, addressing memory stagnation and memory bloat through a dynamic memory management framework. It introduces a Bayesian-inspired memory update guided by memory entropy, a confidence-weighted memory unit structure, entropy-driven compression, and a two-stage retrieval pipeline, all orchestrated by a Master Agent to minimize global belief entropy. The authors validate their approach on the DABench dataset, showing improved personalization, logical coherence, and emotional resonance while substantially reducing memory growth. The proposed DAM-LLM framework and DABench benchmark offer a novel direction for designing memory-centric, emotionally intelligent AI agents with scalable retrieval and robust forgetting.

Abstract

Advances in large language models are making personalized AI agents a new research focus. While current agent systems primarily rely on personalized external memory databases to deliver customized experiences, they face challenges such as memory redundancy, memory staleness, and poor memory-context integration, largely due to the lack of effective memory updates during interaction. To tackle these issues, we propose a new memory management system designed for affective scenarios. Our approach employs a Bayesian-inspired memory update algorithm with the concept of memory entropy, enabling the agent to autonomously maintain a dynamically updated memory vector database by minimizing global entropy to provide more personalized services. To better evaluate the system's effectiveness in this context, we propose DABench, a benchmark focusing on emotional expression and emotional change toward objects. Experimental results demonstrate that, our system achieves superior performance in personalization, logical coherence, and accuracy. Ablation studies further validate the effectiveness of the Bayesian-inspired update mechanism in alleviating memory bloat. Our work offers new insights into the design of long-term memory systems.

Figures (7)

• Figure 1: The DAM-LLM framework: dynamic management of Memory Units via entropy minimization by a Master Agent.
• Figure 2: The collaborative workflow in this work: a question-answering pipeline featuring routing, extraction, and master agents built upon long-term dynamic affective memory—with distinct colored arrows delineating its various processing paths.
• Figure 3: Illustration of Bayesian-inspired update process schematically.
• Figure 4: Confidence evolution across diverse scenarios.
• Figure 5: LLM sentiment scoring: quantitative response to emotional intensity variation.