Study reveals how memory shapes human-AI relationships over 10 sessions

A longitudinal multimodal study by Sumida et al. (Kyoto University) examined how repeated interactions with a memory-augmented conversational agent evolve into relationships. Twenty-four participants each completed 10 sessions, rating five relational constructs — familiarity, self-disclosure, perceived memory, conversational quality, and enjoyment — after every session. The study identified two key dynamics: first, conversational quality strongly predicts in-the-moment enjoyment but has no cross-session effect. In contrast, perceived memory is relationally conditioned — it depends more on prior relational state than on the system's memory capability — and indirectly boosts later enjoyment via increased self-disclosure.

The second dynamic reveals that relationships are punctuated by discrete turning points — crashes (sudden drops in enjoyment) and surges (sudden rises). These turning points are partially traceable in users' multimodal behavior (e.g., speech, facial expressions). Surges are more behaviorally detectable in real time and their enjoyment boost lasts longer than crashes recover. Some crashes can be forecast from person-specific behavioral drift before they occur. The findings suggest long-term human-AI bonds form through both gradual accumulation and abrupt shifts, offering design implications for adaptive AI that can detect and respond to relational turning points.