RECALLbot: Designing Agentic Memory and Reciprocal Disclosure for Human–Chatbot Relationships

Paper Title

RECALLbot: Designing Agentic Memory and Reciprocal Disclosure for Human–Chatbot Relationships

Publication Info

• Topic area: Human-chatbot relationship development through memory-based reciprocal disclosure.
• Keywords: Social chatbots, agentic memory, reciprocal disclosure, human-chatbot interaction, trust, connectedness, LLMs, user control, memory management, self-disclosure.

Background and Problem

• Problem / challenge: Existing social chatbots lack agentic memories and context-aware reciprocal disclosure strategies, leading to one-sided interactions that fail to establish credible social identities or foster deep human-chatbot relationships.
• Significance: Addressing these gaps is critical for enhancing companionship, trust, and emotional support in human-chatbot interactions, which are increasingly used to reduce loneliness and build social connectedness.
• Motivation and related work: Prior work has focused on user-related memories and scripted disclosures but has not equipped chatbots with life-like, agentic memories or dynamic, in-situ disclosure strategies. This paper builds on these findings to integrate memory construction and reciprocal disclosure mechanisms.

Solution

• Proposed approach: RECALLbot, a social chatbot that constructs agentic Me and We Memories and employs context-aware reciprocal disclosure strategies with user controls.
• Novelty:
  1. Introduction of agentic memory design, including life-like Me Memories and co-constructed We Memories, to enhance chatbot authenticity and anthropomorphism.
  2. Development of context-aware reciprocal disclosure strategies that adapt to conversational pacing and user intent, operationalizing interpersonal communication theories.
  3. Empirical evaluation of memory-based reciprocal disclosure’s impact on human-chatbot relationship formation, providing new insights into relational dynamics.
• Procedure and key techniques:
  • Memory Construct Module: Generates Me Memories (background and diary) and We Memories (dialogue summaries and user profiles) using LLMs, with user controls for viewing, editing, and pinning.
  • Memory Disclosure Module: Dynamically plans disclosure turns based on conversational dominance, retrieves topically relevant memories, and generates empathetic, context-aware messages. Includes lightweight controls for pacing, topic selection, and feedback.
  • Implementation: Built using Doubao APIs for LLMs, with memory stored in JSON format and categorized into short-term and long-term memory.

Results

• Concrete findings:
  • RECALLbot elicited 61.26% disclosure frequency compared to 40.11% in the baseline.
  • Users in the RECALLbot group disclosed at deeper levels (51.70% at Level 3) compared to the baseline (16.19% at Level 3).
  • Narrative coherence in chatbot self-disclosure was higher in RECALLbot (66.68%) than in the baseline (55.49%).
  • Trust in the chatbot improved significantly over time, particularly in perceived risk reduction.
• Advantage over baselines: RECALLbot outperformed the baseline system in enhancing perceptions of social identity, eliciting deeper and more frequent disclosures, and fostering trust and engagement.
• Experiments / evaluation: Conducted a two-week between-subjects study with 40 participants (20 per group). Participants interacted with their assigned chatbot for at least 10 minutes daily, with surveys, linguistic analysis of chat logs, and focus group interviews used for evaluation.
• Limitations and future work:
  • Short study duration limited long-term relationship analysis.
  • Culturally homogeneous and gender-skewed participant pool.
  • Text-only interaction excluded multi-modal communication.
  • Persona and disclosure strategies were not fully integrated.
  • Future work should explore longer-term deployments, diverse populations, multi-modal interactions, and multi-agent settings.

Summary

RECALLbot introduces agentic memory and reciprocal disclosure mechanisms to enhance human-chatbot relationships. By constructing Me and We Memories and dynamically adapting disclosure strategies, the system fosters perceptions of social identity, elicits deeper self-disclosures, and builds trust. A two-week user study demonstrated significant improvements in relational outcomes compared to a baseline system. Future research should address long-term dynamics, diverse user needs, and multi-modal communication to further advance memory-enabled social chatbots.