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

This study investigates how repeated interactions with a conversational AI evolve into a relationship. The researchers conducted a longitudinal study with 24 participants interacting with a memory-augmented agent (InteLLA) over 10 daily sessions. By collecting post-session ratings on five key constructs—familiarity, self-disclosure, perceived memory, conversational quality, and enjoyment—the team analyzed both the gradual accumulation of relational growth and the impact of discrete, abrupt turning points known as crashes and surges.

The researchers found that relational growth is not a monolithic process. Conversational quality is a strong predictor of immediate enjoyment but fails to carry over to future sessions. In contrast, perceived memory functions as a longitudinal anchor; it is not merely a technical capability but a relationally conditioned perception that predicts increased self-disclosure in the next session. This self-disclosure, in turn, mediates the link between memory and future enjoyment, suggesting that users build relationships by feeling recognized over time.

Relationships were marked by "crashes" (sharp declines in relational constructs) and "surges" (sharp improvements). The study reveals a critical asymmetry in these events: surges are more detectable through real-time multimodal behavioral cues (audio, video, and text), whereas crashes are often better forecasted by observing behavioral drift in previous sessions. This suggests that AI systems should employ different strategies for these events: monitoring for long-term behavioral decay to prevent crashes, while using real-time feedback to reinforce positive surges.

Understanding that human-AI relationships are built on both slow, memory-driven accumulation and sudden, volatile turning points is essential for designing more resilient conversational agents. By identifying that different relational dimensions serve different temporal functions, developers can move beyond optimizing for single-session "conversational quality" and instead design for long-term relational continuity and proactive intervention.