Skip to content

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

🕒 Published (v1): 2026-07-16 05:43 UTC · Source: Arxiv · link

Why this paper was selected

Longitudinal memory-augmented conversational agent; empirical relational dynamics data

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

A 24-participant × 10-session longitudinal study of a memory-augmented voice agent (InteLLA) decomposes human-AI relational dynamics into two mechanisms: slow cross-session accumulation driven by Perceived Memory and self-disclosure, and abrupt "crash/surge" turning points that exhibit asymmetric detectability from multimodal behavioral signals.

Problem

Nearly all prior human-AI relational research is single-session, leaving unanswered how repeated interactions cohere into an ongoing relationship. Existing longitudinal work relies on descriptive questionnaire analyses without modeling temporal dependencies among relational constructs or characterizing the behavioral signatures of abrupt positive and negative relational shifts.

Method

N=24 university students interacted with InteLLA (GPT-4o-mini + RAG memory, ~5 min/session, 20–25 turns) across 10 daily voice-based sessions. After each session, participants rated five constructs (Familiarity, Social Penetration/self-disclosure, Perceived Memory, Conversational Quality, Enjoyment) on 7-point Likert scales. Temporal structure was modeled via: (1) fixed-effects growth curves; (2) same-session concurrent regressions (Eq. 2); (3) fixed-effects cross-lagged panel models (CLPM, Eq. 3) for all 20 directed pairs; (4) a lagged mediation model (Memory\(_t\) → Social Penetration\(_{t+1}\) → Enjoyment\(_{t+1}\)) with cluster bootstrap (\(B=5000\)).

For crash/surge detection, session-to-session deltas \(\Delta w_k^{(t)} = w_k^{(t)} - w_k^{(t-1)}\) crossing \(\pm 1\sigma\) (leave-one-participant-out thresholds) defined binary events. Elastic-Net Logistic Regression (ENLR) was trained on 351 multimodal features (text via GPT-4.1 annotation, audio via eGeMAPSv02/openSMILE, video via OpenFace 2.0) under leave-one-participant-out cross-validation; AUPRC was the primary metric.

Key Contributions

  • Empirical dissociation between within-session quality (Conversational Quality dominates enjoyment but does not carry forward) and cross-session relational growth (Perceived Memory acts as a longitudinal bridge).
  • Demonstration that Perceived Memory is relationally conditioned—predicted by prior Familiarity (\(\beta=0.290\), \(p=.005\)), Enjoyment (\(\beta=0.343\), \(p<.001\)), and Social Penetration (\(\beta=0.268\), \(p=.033\))—rather than being a direct readout of system capability.
  • Operationalization of crash and surge events in longitudinal human-AI interaction and characterization of their asymmetric behavioral detectability.
  • Design prescription: surges should be reinforced in-session; crashes require cross-session behavioral drift monitoring for prevention.

Results

  • Only Social Penetration showed statistically reliable linear growth over sessions (\(\beta=0.081\), \(p=.003\)); the other four constructs showed no reliable trend.
  • Concurrent regression: Conversational Quality is the strongest within-session predictor of Enjoyment (\(\delta=0.430\), \(p<.001\)); no significant cross-lagged paths emerge from it (\(p>.07\)).
  • Cross-lagged paths: Memory\(_t\) → Social Penetration\(_{t+1}\) (\(\beta=0.165\), \(p=.001\)); Enjoyment\(_t\) → Social Penetration\(_{t+1}\) (\(\beta=0.280\), \(p=.022\)).
  • Mediation: Memory → Social Penetration → Enjoyment indirect effect supported via cluster bootstrap; direct Memory → Enjoyment path not significant.
  • Crash rates: 8.9% (Social Penetration) to 17.3% (Familiarity); surge rates: 13.9% (Enjoyment) to 22.8% (Conversational Quality) across 202 transitions.
  • Enjoyment delta distribution is negatively skewed (skew = −0.89, kurtosis = 3.22), confirming crashes are more severe than surges.
  • Surges are more detectable from same-session multimodal behavior than crashes (across all 7 feature-set ablations with ENLR); crashes are better forecast from prior-session temporal/person-normalized features.
  • Enjoyment surges persist more reliably than enjoyment crashes recover.

Limitations

  • Small sample (\(N=24\)), precluding a full Random-Intercept CLPM to formally separate stable trait variance from within-person dynamics.
  • Heterogeneous recording conditions (personal webcams) introduce measurement noise in video features; video features were restricted to 4 robust measures.
  • All participants were university students with CEFR C1+ English; generalizability to broader populations is unknown.
  • Sessions were short (~5 minutes), limiting discourse depth and potentially compressing the relational variance observable per session.
  • The CLPM specification is a short-panel lagged regression, not a full RI-CLPM, which may conflate between-person and within-person effects.
  • RAG memory performance was not independently measured; Perceived Memory reflects user experience, not verified system recall accuracy.

Relevance to Agentic AI / LLM Agents

This study directly informs the design of memory-augmented LLM agents intended for repeated, longitudinal use—the dominant deployment paradigm for personal AI companions and persistent assistants. The finding that Perceived Memory is relationally conditioned (not a direct capability readout) has immediate implications for how RAG-based memory retrieval should be calibrated and surfaced: raw recall accuracy is insufficient; relational state must be considered. The crash/surge asymmetry framework and ENLR-based behavioral monitoring pipeline provide a concrete architecture for adaptive agents that distinguish in-session surge reinforcement from cross-session crash prevention—directly relevant to RLHF-based agent fine-tuning and real-time intervention systems.