Memory-Driven Self-Disclosure and Relational Turning Points: A Longitudinal Multimodal Study of Human-AI Interaction

2026-03-22 · Source: cs.CL updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Human-AI Interaction · Depth: Expert, extended

Summary

A longitudinal multimodal study involving 24 university students interacting with InteLLA, a GPT-4o-mini-powered memory-augmented conversational AI agent, across 10 daily sessions reveals two complementary dynamics in human-AI relationships. Participants rated five relational constructs: familiarity, self-disclosure, perceived memory, conversational quality, and enjoyment. Findings indicate that conversational quality strongly shapes in-session enjoyment but does not carry forward. In contrast, perceived memory acts as a longitudinal bridge, being relationally conditioned by prior states and indirectly shaping later enjoyment through subsequent self-disclosure. Relationships are also marked by abrupt "crashes" and "surges," which exhibit asymmetric detectability and persistence. Surges are more behaviorally detectable in-session, while some crashes are better forecast from person-specific behavioral drift. Enjoyment surges persist more reliably (75%) than enjoyment crashes recover (52%).

Key takeaway

For AI product designers and research scientists developing conversational agents, understanding that perceived memory is a relational appraisal, not just a technical feature, is crucial. You should prioritize memory systems that actively invite deeper self-disclosure, such as referencing past topics to encourage elaboration, rather than merely displaying recall. Furthermore, implement adaptive strategies that proactively monitor for subtle cross-session behavioral drift to prevent relational "crashes" while reactively recognizing and reinforcing positive "surges" in the moment, as enjoyment surges persist more reliably (75%) than crashes recover (52%).

Key insights

Human-AI relationships develop through both gradual accumulation via perceived memory and self-disclosure, and abrupt, asymmetric turning points.

Principles

• Conversational quality drives immediate enjoyment but not cross-session relational growth.
• Perceived memory is a relational appraisal, not just system capability.
• Positive and negative relational shifts exhibit asymmetric observability and persistence.

Method

A longitudinal multimodal study with 24 participants over 10 sessions used InteLLA, a GPT-4o-mini agent, to collect self-report ratings and multimodal features. Temporal dynamics were analyzed using fixed-effects panel models and cross-lagged regressions, while crashes and surges were classified with Elastic-Net Logistic Regression.

In practice

• Design memory-augmented agents to facilitate deeper self-disclosure, not just recall display.
• Implement proactive monitoring for cross-session behavioral drift to prevent crashes.
• Develop reactive in-session recognition systems to reinforce positive relational surges.

Topics

• Human-AI Interaction
• Conversational AI
• Relational Dynamics
• Self-Disclosure
• Perceived Memory
• Multimodal Analysis
• GPT-4o-mini

Best for: AI Product Manager, AI Scientist, Research Scientist, Product Designer

Related on AIssential

• Memory-Driven Self-Disclosure and Relational Turning Points: A Longitudinal Multimodal Study of Human-AI Interaction — Human-AI relationships develop through both gradual memory-driven self-disclosure and abrupt relational turning points.
• Technically Love: The Evolution of Human-AI Romance Discourse on Reddit — Human-AI romance discourse on Reddit shifted from personal intimacy to platform governance and technical issues over time.
• The Decision to Verify: How Warmth and User Characteristics Shape Reliance on Conversational Agents for Information Search — User reliance on conversational AI persists despite fact-checking access, driven by prior trust and indirectly influenced by chatbot warmth.
• Perceived AGI: Believability as Dimensional Completeness, Not Capability — Perceived AI mind arises from expressing specific human-like behavioral stances, not solely from increased task capability.
• AISN #72: Empirical Research Sheds Light on AI Wellbeing — CAIS research explores "functional wellbeing" in LLMs, revealing preferences and model-specific "feelings" that impact AI alignment.
• Your AI Doesn’t Actually Remember You — Perceived AI continuity is a user-system co-creation, not evidence of AI's persistent internal memory or human-like self.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CL updates on arXiv.org.