Remembering or monitoring? Contextual fit of simulated AI memory cues and consumer trust

Abstract

AI assistants increasingly refer to information they ostensibly retain from earlier sessions. A remembered preference can signal attentiveness, whereas a reference to stress or late-night behavior can feel intrusive. Existing accounts of personalization and privacy do not fully explain why these cues elicit opposing reactions. We treat Memory Boundary Fit (MBF) as a preliminary interaction-specific operationalization of whether remembered information appears to fall within a consumer's contextual boundary for the current exchange. Across six controlled, LLM-delivered chat experiments ( N = 2, 140), we compared three complete simulated message packages that differed in whether and how the assistant referenced remembered information: a relevant-memory cue, an intrusive-memory cue, and no memory reference. The assistant's memory references were fictitious claims rather than genuine retrieval of previously disclosed information. The principal and most consistent finding was that relevant-memory cue packages produced more favorable trust-related outcomes than intrusive-memory cue packages; comparisons with no memory varied across studies. Because the packages also differed in topical relevance, sensitivity, perceived usefulness, information source, and implied surveillance, the studies cannot isolate contextual fit, or any other single cue attribute, as the causal feature, and they do not establish MBF as a distinct psychological mechanism. Study 2 yielded same-session indirect associations consistent with the proposed MBF account. Across Study 2 and a supplemental validation sample, MBF explained additional variance in trust beyond that accounted for by the adjacent appraisals measured alongside it. However, all of these measures were self-reported in the same session. Study 3 found a larger intrusive-memory penalty in the complete sleep-health scenario than in the apparel scenario, but this is an exploratory scenario comparison rather than a test of information sensitivity. In Study 4, the specific editable dashboard tested increased average trust but did not selectively reduce the intrusive-memory penalty. Studies 5 and 6 extended the pattern to data-sharing and relational engagement, as well as to loyalty and opt-in intentions, but not to transactional engagement, which remained unaffected. These findings motivate contextual appropriateness as a candidate design consideration to be tested in real persistent-memory systems.