Commonsense-augmented Memory Construction and Management in Long-term Conversations via Context-aware Persona Refinement

TL;DR

The paper tackles the problem of uninformative personas in long-term conversations by introducing Caffeine, a framework that performs commonsense-based persona expansion via COMET and then refines contradictory personas through context-aware strategies in a graph-based iterative process. It combines memory retrieval with LLM-driven refinement (Resolution, Disambiguation, Preservation) to produce richer, context-grounded persona sentences stored in long-term memory for future sessions. Empirical results on Multi-Session Chat show improved response generation, enhanced alignment with human preferences, and significant cost/time efficiency gains compared to baselines. Overall, the work demonstrates that context-aware refinement of contradictory personas yields more informative speaker representations and more engaging, coherent long-term dialogue capabilities.

Abstract

Memorizing and utilizing speakers' personas is a common practice for response generation in long-term conversations. Yet, human-authored datasets often provide uninformative persona sentences that hinder response quality. This paper presents a novel framework that leverages commonsense-based persona expansion to address such issues in long-term conversation. While prior work focuses on not producing personas that contradict others, we focus on transforming contradictory personas into sentences that contain rich speaker information, by refining them based on their contextual backgrounds with designed strategies. As the pioneer of persona expansion in multi-session settings, our framework facilitates better response generation via human-like persona refinement. The supplementary video of our work is available at https://caffeine-15bbf.web.app/.

Figures (12)

• Figure 1: Contradictory personas can co-exist and provide rich speaker information for the conversation when their contexts are considered (an empirical example).
• Figure 2: At the end of each dialogue session, Caffeine refines contradictory personas within/across the session(s) and saves the refined version to the dialogue model's memory for response generation in the next session.
• Figure 3: Empirical demonstration of our strategies. Top: relevant contexts; Mid: contradictory personas; Bottom: refined persona(s).
• Figure 4: Human evaluation results on (i) refined personas and (ii) the refinement process (p-value $< 0.05$).
• Figure 5: Cost and time efficiency of our algorithm.