Measure usefulness, perceived understanding, trust and anthropomorphism separately.
A remembered detail can do two different things
Suppose an assistant remembers that a user prefers concise explanations. On the next visit it answers more briefly. That is functional personalization. Now suppose it remembers a difficult conversation and refers to it in an emotionally resonant way. The same technical capability—retrieving information from earlier interactions—may be interpreted as attention, continuity or care. I want to understand where this shift in interpretation begins.
My research question is not whether AI systems are people or possess feelings. It is whether particular kinds of persistent memory cause users to attribute more human-like social understanding to an assistant, even when the assistant's underlying capabilities remain unchanged.
Anthropomorphism is not the same as usefulness
An assistant can be more useful without appearing more human. It can also appear more attentive without becoming more accurate. That is why a study should measure several outcomes separately: task usefulness, perceived understanding, trust, anthropomorphism, emotional closeness, willingness to disclose and perceived control over personal information.
These constructs can move in different directions. Episodic memory might make an assistant feel attentive while simultaneously making some users uncomfortable. A factual reminder might improve task performance without increasing relational attachment. Treating all positive reactions as 'trust' would erase these distinctions.
A proposed controlled comparison
I would compare four memory conditions using the same underlying model, task instructions and interaction opportunities. The first condition has no cross-session memory. The second remembers stable user facts relevant to a task. The third remembers expressed preferences, such as explanation style. The fourth remembers episodic or relational context from earlier conversations. Each condition would disclose memory use in a consistent, ethically appropriate manner.
The study could use repeated interactions rather than a single impression survey, because continuity is itself part of the mechanism being examined. Outcomes would be measured after matched tasks and at later intervals. Assignment should be randomized where possible, and task performance should be evaluated independently of subjective impressions.
The provenance question
Memory also introduces an epistemic problem: how does the assistant know what it claims to remember? A statement may come from an explicit user message, an inference, a summary of an older interaction or a third-party source. Presenting all of these as equally certain recollections risks creating a false sense of intimacy and accuracy.
This connects to my Memory-Provenance Audit research proposal. One possible evaluation protocol would vary whether an assistant attributes a remembered detail to a direct statement, an inferred preference or an uncertain summary, then measure how users interpret confidence, accuracy and agency. A responsible system should make it possible to correct or reject a memory rather than allowing an unverified inference to harden into an apparent fact.
Possible outcomes and counter-hypotheses
One hypothesis is that episodic memory increases perceived social understanding more strongly than factual memory, even when factual memory is equally useful for task completion. A competing hypothesis is that explicit provenance labels reduce anthropomorphic interpretation by reminding users that the system is retrieving records rather than personally remembering. Another possibility is that transparency improves trust even while reducing perceived human-likeness.
These possibilities are deliberately open. The effect may depend on age, familiarity with AI, emotional context, task stakes and the user's existing expectations. An ethical study would avoid deliberately encouraging dependency and would provide clear consent and control over what information is retained.
Why the distinction matters
Persistent memory may make AI tools genuinely more helpful, especially for complex tasks that span weeks or months. But design choices about what is remembered, how it is surfaced and how uncertainty is communicated also shape the relationship people believe they have with the system.
I see this as a human–AI interaction research agenda rather than a completed empirical finding. The goal is to understand how technical memory architectures influence human interpretation—and how systems might preserve the benefits of continuity without manufacturing unwarranted social authority.