Modeling Long-Term Memory for Multi-Session Task-Oriented Dialogue Systems via Memory-Active Policy

ACL ARR 2025 February Submission3824 Authors

Abstract: Existing Task-Oriented Dialogue (TOD) systems generally focus on single-session dialogues and overlook the study of multi-session interactions, leading to the inability to track long-term memory to obtain target-related information from previous dialogue sessions for more efficiently personalized interaction in TOD. To address this challenge, we introduce a \textbf{MS-TOD} dataset, the first multi-session TOD dataset designed to retain long-term memory across sessions, enabling fewer turns and more efficient task completion. Based on this new dataset, we propose a \textbf{Memory-Active Policy (MAP)} that improves multi-session dialogue efficiency by reducing turns through a two-stage approach. Specifically, we first introduce Memory-Guided Dialogue Planning, which retrieves relevant history through intent descriptions, utilizes a memory judger to identify key QA units, and employs a reader to generate responses based on reconstructed memory. Next, the Proactive Response Strategy is designed to detect and correct errors or omissions, ensuring efficient and accurate task completion. We evaluate MAP on our MS-TOD dataset, focusing on response quality and effectiveness of the proactive strategy. Experimental results show that MAP enhances multi-session TOD performance by improving turn efficiency and task success through long-term memory integration while maintaining comparable performance in single-session multi-turn tasks.

Paper Type: Long

Research Area: Dialogue and Interactive Systems

Research Area Keywords: task-oriented dialogue, multi-session, long-term memory

Contribution Types: NLP engineering experiment, Data resources

Languages Studied: English

Submission Number: 3824