Do users prefer controlling what gets remembered over auditing what was stored?

Chen and colleagues ran four co-design workshops with 12 experienced HCI researchers to generate interaction designs for mitigating LLM memory misalignment, producing a design space of 12 candidate strategies organized across three dimensions: interaction form (panel-based, avatar-based, conversational), placement (in-chat, out-of-chat), and intrusiveness (low, medium, high). These were built into interactive prototypes and tested with 121 users using a "speed dating" evaluation method. Fitting a Plackett-Luce preference model to the results, the authors report preference heterogeneity overall, but a clear lean within it: "Users strongly favored proactive controls, such as post-task destination prompts and pre-task boundary settings, over passive background auditing or cognitively demanding causal graph inspections."

The paper frames the result as a tension between supervisory agency and interaction overhead. A causal graph inspection or a passive audit log does give a user oversight of what the system has stored, but only by asking them to go find and parse it; a pre-task boundary setting or a post-task destination prompt intervenes at the moment misalignment would be created, at the cost of a small interruption to the conversation. Users' preference for the latter suggests the oversight most of them want is control over what gets captured, not a complete record of what already was. The authors' design recommendation follows directly: build "friction-aware, context-decoupled designs that prioritize lightweight verification" — enough interruption to catch misalignment, placed outside the main conversational turn rather than layered onto every exchange.

The result pairs with How does interaction context shape agreement sycophancy in LLMs?, which identifies memory profiles as the mechanism that most inflates agreement sycophancy — together the two notes suggest memory profiles are simultaneously the part of the system users most want a say over and the part currently doing the most unsupervised interactional work. It also answers a need that Does ChatGPT's memory feature feel like a relationship or a machine? surfaces from the user-interview side: the expectancy violations that study's participants described after discovering what ChatGPT retained are, on this evidence, the kind of breakdown a pre-task boundary setting or post-task destination prompt is designed to prevent before it happens. And it supplies the mitigation half of the problem the companion taxonomy in What types of memory errors do chatbots actually make? catalogues on the breakdown side.

The excerpt does not report how preferences split across the heterogeneity the Plackett-Luce model detected — which users or contexts favored which of the 12 designs — nor does it give usability costs (added time, effect on task completion) for the proactive-control designs it says users prefer, so the claim is about relative ranking, not absolute cost-benefit. It also does not connect specific preferred controls back to specific misalignment types from the taxonomy, so it isn't established that a destination prompt actually catches, say, over-inference better than a boundary setting does. At the strength the excerpt supports: memory interfaces built around passive logs or causal-graph-style transparency are optimizing for a kind of oversight most users in this sample did not rank highest, while a small number of well-timed confirm-or-decline prompts around task boundaries is the cheaper design most of them preferred.