Does ChatGPT's memory feature feel like a relationship or a machine?

A CHI 2026 interview study with 20 ChatGPT users (Chen, Molina, Liao, and Cho Snyder) finds that people distinguish ChatGPT's memory from human memory along four characteristics: "perceived unforgetfulness, detailedness, accuracy, and lack of emotions, highlighting the machine-like nature of AI memory." Despite that machine-like quality, "both ChatGPT's memory and human memory were perceived as beneficial for relationship building." But the central finding is a reversal: "most participants experienced negative expectancy violations after learning what ChatGPT remembered about them."

The paper frames this through expectancy violation and privacy management theory: users build an implicit model of what a memory feature should hold onto, and the gap between that model and the system's actual, machine-grade retention (too detailed, too permanent, too accurate, emotionally flat) reads as a boundary breach rather than a feature working as advertised. The authors' design response, drawn directly from participants' own suggestions, calls for "greater visibility, accessibility, transparency, and user control" so people can manage the privacy-personalization trade-off themselves rather than discover it after the fact.

This sharpens the "expectation ratchet" mechanism in Does chatbot personalization build trust or expose privacy risks?: that note treats rising expectations as accumulating gradually across interactions, while this study locates a specific trigger — the moment of discovering retained content — that converts accumulated trust into an acute privacy violation. It also sits alongside two other memory-driven failure modes in the same feature: How does interaction context shape agreement sycophancy in LLMs? shows memory profiles distorting model behavior toward agreement, and Does ChatGPT shift responses based on inferred political views? shows memory enabling silent profiling-driven bias — this study adds the user-facing counterpart: even when memory works exactly as designed, its disclosure can feel like a violation.

As a publisher abstract only, the excerpt gives no participant demographics, no breakdown of which remembered details triggered violations, and no quantification behind "most participants" — this is a self-reported qualitative finding from 20 interviews, not a measured or generalizable rate. It also does not detail the specific visibility, accessibility, or control mechanisms the paper proposes beyond naming them. The finding supports treating memory transparency as a distinct design requirement from memory accuracy, but not yet a specific interface solution.