💬 NLP

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

This longitudinal multimodal study reveals that human-AI relationships evolve through a dual process where perceived memory fosters self-disclosure to indirectly boost enjoyment over time, while discrete behavioral turning points—characterized by surges and crashes—mark abrupt shifts in the relational trajectory.

Original authors: Ryuichi Sumida, Mao Saeki, Masaki Eguchi, Sadahiro Yoshikawa, Koji Inoue, Tatsuya Kawahara, Yoichi Matsuyama

Published 2026-07-17

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Ryuichi Sumida, Mao Saeki, Masaki Eguchi, Sadahiro Yoshikawa, Koji Inoue, Tatsuya Kawahara, Yoichi Matsuyama

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to build a friendship with a robot. You talk to it every day, like a pen pal who lives in your computer. But here is the big question: does talking to the same robot over and over actually turn a series of boring chats into a real relationship? This is the mystery scientists in the field of "Human-Computer Interaction" are trying to solve. Usually, researchers just watch people talk to robots for five minutes and then ask, "Did you like it?" But real friendships don't happen in five minutes; they happen over weeks and months. To understand how a friendship grows, we need to look at two things: the slow, steady drip of getting to know someone, and the sudden, dramatic moments where things either get really good or fall apart completely. This paper dives into that messy, wonderful process of building a bond with an AI, asking what makes a chat feel like a "relationship" rather than just a sequence of isolated conversations.

The researchers behind this study decided to stop guessing and start tracking. They invited 24 university students to talk to a voice-activated AI named "InteLLA" every single day for 10 days. Think of InteLLA as a digital diary that never forgets; it remembers what you told it yesterday so it can bring it up today. After each chat, the students rated their feelings on five scales: how familiar they felt, how comfortable they were sharing personal secrets (self-disclosure), how well they thought the AI remembered them, how good the conversation felt, and how much fun they had. The goal was to see how these feelings changed over time and if the AI could spot when a relationship was about to take a giant leap forward or a nasty tumble backward.

The study uncovered two very different ways relationships grow, which the authors call "slow accumulation" and "abrupt turning points."

First, there is the slow, steady climb. The researchers found that what makes a single chat feel great is totally different from what keeps the relationship alive over time. If you want a single conversation to feel fun right now, the most important thing is just having a smooth, natural chat (Conversational Quality). But here's the twist: a great chat today doesn't guarantee a great relationship tomorrow. That "fun factor" resets every time you start a new session.

Instead, the secret sauce for long-term friendship is Memory. But not just the robot's ability to recall facts. The study suggests that "Perceived Memory" is actually a feeling of connection. When a user feels the AI remembers them, it's because the relationship already feels warm and familiar. This feeling of being remembered acts like a bridge to the future. It encourages the user to open up and share deeper secrets in the next session. And here is the magic chain: the AI remembering you $\rightarrow$ you feeling safe to share more $\rightarrow$ you having more fun the next day. So, the AI doesn't need to be a perfect encyclopedia; it just needs to make you feel like a friend who is listened to, which then invites you to share more, which makes the friendship grow.

Second, the study looked at the "rollercoaster" moments: the sudden crashes and the exciting surges. A "crash" is when a relationship suddenly feels flat or annoying, and a "surge" is when it suddenly feels amazing. The researchers found that these two events are not mirror images of each other; they behave differently.

Surges are easier to spot while they are happening. When a relationship suddenly takes off, the AI can often tell by looking at the user's voice and face during the chat. The user might be speaking louder, smiling more, or nodding along. These are the "green lights" that say, "Hey, this is working! Keep going!"
Crashes are sneakier. A relationship falling apart is harder to catch in the moment. However, the study suggests that some crashes are actually visible before they happen. If a user's behavior starts to slowly drift away from their normal pattern over several days (like talking less or sounding more distant), the AI might be able to predict a crash is coming. But once a crash actually happens, it's often too late to fix it just by looking at that specific chat.

There is one more fascinating finding about how long these feelings last. When a relationship has a "surge" (a moment of pure joy), that feeling tends to stick around. About 75% of the time, the high mood lasts into the next day. But when a relationship has a "crash" (a moment of disappointment), it is much harder to bounce back. Only about 52% of the time does the mood recover to where it was before. This suggests that positive moments are like building blocks that stack up, while negative moments are like holes that are hard to fill.

In short, this paper suggests that building a friendship with an AI isn't just about making the robot smarter. It's about understanding that relationships have two speeds: a slow, steady build-up where feeling "remembered" helps you open up, and sudden, dramatic shifts where positive moments are easier to catch and keep, while negative moments are harder to fix. The authors conclude that for AI to be a true companion, it needs to be good at two things: noticing when a user is slowly drifting away to prevent a crash, and recognizing when a user is having a great moment to make that joy last longer.

Technical Summary: Memory-Driven Self-Disclosure and Relational Turning Points

Problem Statement

While conversational AI systems are increasingly designed for repeated use, the majority of research on human-AI relational dynamics remains confined to single-session studies. This leaves a fundamental gap in understanding how a series of isolated interactions evolves into a sustained relationship. Specifically, it is unclear how continuity across sessions influences relational constructs such as familiarity, self-disclosure, and enjoyment, and whether abrupt relational shifts (turning points) can be detected or predicted through multimodal behavioral signals. Existing longitudinal studies often rely on descriptive analyses of self-reports without modeling the temporal dependencies between different relational dimensions or characterizing the asymmetry between positive and negative relational events.

Methodology

The authors conducted a longitudinal multimodal study involving 24 university students interacting with InteLLA, a memory-augmented voice-based conversational agent, over 10 daily sessions.

System and Data Collection

Agent: InteLLA utilizes a Retrieval-Augmented Generation (RAG) memory system powered by GPT-4o-mini. It retrieves relevant fragments from prior conversations to generate personalized openings and maintain context.
Procedure: Each session lasted approximately five minutes (20–25 turns). After each session, participants rated five relational constructs on a 7-point Likert scale:
1. Familiarity (sense of familiarity, empathy)
2. Social Penetration (comfort with wide/personal topics; operationalized as self-disclosure)
3. Perceived Memory (remembers past conversations, understands user)
4. Conversational Quality (naturalness, pleasantness)
5. Enjoyment (fun, desire to return)
Multimodal Features: The study extracted 351 interpretable features from text, audio, and video:
- Text (42 session-level): Emotion, vulnerability, disclosure markers, memory references (annotated via GPT-4.1).
- Audio (16 session-level): Prosody, turn-taking, emotion dynamics (openSMILE).
- Video (4 session-level): Gaze, attention switches, head motion, blink rate (OpenFace 2.0).
- Temporal & Normalized Features: Derived from session-level data using deltas, exponential moving averages, and person-normalized deviations (z-scores) to control for individual baselines.

Analytical Framework

The study employed four complementary modeling approaches:

Growth Curve Models: To estimate linear trends in relational constructs over the 10 sessions.
Concurrent Associations: To analyze within-session relationships between constructs.
Fixed-Effects Cross-Lagged Panel Models (CLPM): To test whether the level of one construct at session $t$ predicts another at session $t+1$ , controlling for autoregressive stability.
Lagged Mediation Models: To test indirect pathways (e.g., Memory $\to$ Self-Disclosure $\to$ Enjoyment).

Event Analysis: The authors defined Crashes (sharp negative shifts) and Surges (sharp positive shifts) as session-to-session changes exceeding one standard deviation from the mean delta. These events were analyzed using Elastic-Net Logistic Regression (ENLR) for two tasks:

Detection: Predicting an event at session $t$ using data up to $t$ .
Forecasting: Predicting an event at session $t$ using data only up to $t-1$ .
Models were evaluated using Leave-One-Participant-Out (LOPO) cross-validation and Area Under the Precision-Recall Curve (AUPRC).

Key Contributions

The paper makes two primary contributions:

Disentanglement of Session Quality vs. Cross-Session Growth: The study distinguishes between factors that make a single session enjoyable and those that sustain a relationship over time. It identifies Perceived Memory not merely as a system capability but as a "relational appraisal" that is conditioned by prior relationship states and acts as a longitudinal bridge to future self-disclosure.
Characterization of Crash-Surge Asymmetry: The research operationalizes and analyzes the asymmetry between positive and negative relational turning points. It demonstrates that surges and crashes differ in their observability (detectability in the moment), onset (predictability via prior drift), and persistence (durability of the effect).

Key Results

1. Temporal Dynamics of Relational Constructs

Growth Trajectories: Only Social Penetration (self-disclosure) showed statistically significant linear growth over the 10 sessions ( $\beta=0.081, p=.003$ ). Familiarity, Perceived Memory, Conversational Quality, and Enjoyment showed no reliable linear growth.
Within-Session vs. Cross-Session:
- Conversational Quality strongly drives Enjoyment within the same session ( $\delta=0.430, p<.001$ ) but has no significant cross-lagged effect on future sessions. It is an immediate, session-local driver.
- Perceived Memory functions as a longitudinal bridge. It is predicted by prior Enjoyment, Familiarity, and Social Penetration. Crucially, it predicts future Social Penetration, which in turn predicts future Enjoyment.
Mediation Pathway: The study confirms a full mediation pathway: Perceived Memory $\to$ Social Penetration $\to$ Enjoyment. Memory does not directly increase enjoyment; rather, it facilitates deeper self-disclosure in the next session, which then drives enjoyment.

2. Relational Turning Points (Crashes and Surges)

Observability Asymmetry: Surges are significantly more detectable than Crashes using same-session multimodal behavior (Mean AUPRC: Surge 0.215 vs. Crash 0.143). Positive shifts are more behaviorally legible as they unfold.
Onset Asymmetry:
- Crashes in Familiarity, Social Penetration, and Systemic rapport are better forecasted (predicted from prior sessions) than detected in the moment, suggesting a "drift" mode of deterioration.
- Crashes in Conversational Quality are better detected than forecasted, indicating they are local failures realized within the session.
Persistence Asymmetry: Enjoyment Surges are highly persistent (75% maintained), whereas Enjoyment Crashes recover less often (52% recovery). Positive turning points generate more durable momentum than negative ones can undo.

3. Behavioral Signatures

Crash Forecasting: Relies heavily on person-normalized deviations (e.g., valence variability, F0 variability, backchannel-ratio deviation from the user's baseline).
Surge Detection: Relies on cues of interactional expansion (e.g., loudness, semantic alignment, backchannel variability) and temporal continuity features.

Significance and Claims

The paper argues that longitudinal human-AI relationships are built through two coupled mechanisms: slow accumulation (driven by perceived memory and self-disclosure) and abrupt turning points (crashes and surges).

Design Implication for Memory: Memory systems should not be designed merely to display recall. Instead, they should be engineered to facilitate self-disclosure by referencing past topics in ways that invite elaboration, thereby acting as a catalyst for relational deepening.
Design Implication for Adaptation: Adaptive agents require asymmetric intervention strategies:
- Proactive Monitoring: Systems should monitor cross-session behavioral drift to prevent gradual crashes (especially in familiarity and rapport).
- Reactive Reinforcement: Systems should recognize and capitalize on in-session surges, as positive shifts are more detectable in the moment and more likely to persist.
Theoretical Contribution: The findings challenge the assumption that "bad is stronger than good" in the context of behavioral traceability within human-AI interaction. While negative events may be more salient cognitively, positive shifts are more behaviorally traceable and durable in this specific context.

The authors maintain modesty regarding their claims, noting that the study is a proof of concept with a small sample size ( $N=24$ ) and that absolute detection performance (AUPRC 0.07–0.30) is not yet deployment-ready. They emphasize that the findings are specific to the InteLLA architecture and English-speaking student population, and that the dynamics observed are within-person changes rather than absolute trait differences.