A new study says AI assistants have been personalizing you wrong.
Researchers built something they call the Atomic User Model, a structured profile of a person's psychology, thinking style, experience, behavior, and social traits, organized around a stable "identity core." Instead of stuffing a full write-up into every prompt, the system runs a task classifier that pulls only the handful of profile fields relevant to what you're actually asking for. In tests across 16 simulated users and six writing tasks, eight retrieved fields matched the quality of the full profile while using just 23 percent of the tokens - 211 instead of 915. The system also got much better at picking out a user's own writing from four samples, jumping from 14.9 percent accuracy to 42.7 percent, against a 25 percent baseline for random guessing.
Most personalization today amounts to a preference note stapled to every prompt - short answers, no emoji, that sort of thing - which treats voice as a checklist instead of a personality, so the assistant relearns you every time the task changes. The researchers also name a real problem worth knowing about: "personality seepage," where a chatbot picks up traces of your personality just from how you phrase a prompt and imitates it badly, without ever being told who you are. Anchoring memory to a stable identity rather than a shifting preference list is the more useful idea here, and it points at where memory features in tools like ChatGPT and Claude are already headed, just without the formal structure.
Sixteen simulated users and three random seeds is a lab result, not a product - worth watching, not worth rewriting your assistant's memory settings over yet.