Creating an Atomic User Model for Personality-Aware Large Language Model Interaction

Assistants built on large language models are increasingly expected to write as their user would write, and the dominant approach to that expectation operates on a single channel: past preferences are summarised out of conversation history and reinserted into the context window. We argue that this inverts the natural order of inference. Preferences are the task-dependent surface of an underlying personality structure that is comparatively stable, so a system that stores only preferences must relearn the person whenever the task changes. This paper makes four contributions. First, we characterise a phenomenon we call personality seepage, in which the linguistic surface of a prompt carries a personality fingerprint that the assistant mirrors without any access to the personality that produced it. Second, we propose the Atomic User Model (AUM), a structured, human-readable representation that organises a person as a stable identity nucleus surrounded by four interpretable shells covering psychological, cognitive and experiential, behavioural, and social content, together with cross-shell entries recording internal conflict and authenticity.

Introduction. Adaptive systems have modelled their users for as long as they have adapted to them. The generic user modelling tradition in human-computer interaction established that a user model is a reusable component with an explicit schema, maintained separately from the application that consumes it (Kobsa, 2001; Brusilovsky and Millán, 2007), and personalised search built profiles from observed behaviour: interests and activities harvested from a desktop index (Teevan et al., 2005), click and query histories evaluated at scale (Dou et al., 2007), and short-term session signals combined with long-term interest (Bennett et al., 2012). Assistants built on large language models (LLMs) have inherited this design in a compressed form.

Discussion / Conclusion. The finding we would put first is not the effect size but the ratio. Eight retrieved fields, 211 tokens, matched the fidelity of injecting the entire 32-field user model, 915 tokens. Nothing in the study suggests that the remaining twenty-four fields were doing work that the eight did not. For efficiency this is a modest saving. For architecture it is the whole argument. Section 4 committed aum to local residence: the store sits on the user’s device or in a user-controlled vault, never in a third-party log. That commitment is easy to state and hard to keep if the whole model must be transmitted on every complex request, because a model that is transmitted in full on every turn has, for practical purposes, been uploaded. A budgeted payload changes the character of the commitment. If eight fields suffice, disclosure per query is bounded by construction, the bound is measurable, and the fields that never leave are disproportionately the inner-shell ones, which are exactly the ones whose exposure carries the greatest cost. In the vocabulary of contextual integrity (Nissenbaum, 2011), the design question is not whether the user model is secret but whether each flow out of it matches the norms of the context that produced it.