Beyond Memory: Constructing Hierarchical User Identity from Dialogue

ACL ARR 2026 May Submission15438 Authors

Keywords: dialogue systems, user modeling, personalized dialogue, long-term memory, user identity modeling, conversational agents

Abstract: Existing dialogue memory systems mainly optimize storage and retrieval of past utterances, summaries, or extracted facts. For long-term personalization, however, memory should be treated as evidence rather than the final user representation. We formulate hierarchical user identity construction from dialogue, which maps multi-session conversations to a persistent, revisable user state with three layers: factual identity, preference structure, and slow-moving interaction state. We instantiate an inference-first framework that first proposes dialogue-grounded evidence and then applies layer-specific promotion, abstention, and update rules to construct and maintain this state. Across manually reviewed PersonaChat evaluations, this formulation improves explicit fact and preference construction over same-model direct extraction on a larger reviewed overlay, with the same mechanism also visible in a smaller controlled audited subset; a lightweight downstream response-selection pilot further shows that the resulting state improves candidate choice even under a fixed scorer. Reviewed MSC analyses further indicate that the resulting state supports conflict-aware chronological maintenance, while a reviewed adjacent-session analysis suggests that slow-moving interaction state can also be modeled usefully. Taken together, these findings suggest a path beyond memory-centric dialogue systems toward conversational agents that construct and maintain explicit, revisable user identity over time in long-horizon settings.

Paper Type: Long

Research Area: Dialogue and Interactive Systems

Research Area Keywords: dialogue systems, personalization, user modeling, long-term memory, conversational agents, dialogue state tracking

Contribution Types: Model analysis & interpretability, NLP engineering experiment, Data analysis, Position papers, Theory

Languages Studied: English

EMNLP 2026 AI Reviewing Experiment: no

Submission Number: 15438