Dynamic Affective Memory Management for Personalized LLM Agents

2025-10-31

AI Summary

To address emotional memory redundancy, obsolescence, and insufficient contextual integration in personalized AI agents, this paper proposes the Dynamic Emotional Memory System (DEMS). DEMS introduces a Bayesian-inspired memory update mechanism coupled with memory entropy modeling, enabling autonomous memory evolution and compression by minimizing global memory entropy. It constructs a dynamic vector database supporting emotional semantic encoding and context alignment. Additionally, we release DABench—the first benchmark dedicated to emotional expression modeling and longitudinal change assessment. Experiments demonstrate that DEMS significantly improves response personalization, logical coherence, and factual accuracy; ablation studies confirm its effectiveness in mitigating memory bloat. Our core contributions are: (1) an evolvable, entropy-driven memory maintenance paradigm; (2) an emotion-augmented memory–context co-adaptation mechanism; and (3) DABench, a specialized benchmark for evaluating emotional memory capabilities.

Technology Category

• Humans and AI: Emotional Intelligence
• Cognitive Modeling & Cognitive Systems: Affective Computing
• Multiagent Systems: Adversarial Agents

Application Category

• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions
• Search and Retrieval-Augmented AI: Personalized, context-aware and across-device search
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking

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.

Problem

• Addressing memory redundancy and staleness in personalized AI agents
• Improving memory-context integration through dynamic affective memory management
• Developing a benchmark for evaluating emotional expression and change

Innovation

• Bayesian-inspired memory update algorithm for dynamic management
• Memory entropy minimization to reduce redundancy and staleness
• DABench benchmark for evaluating emotional expression performance