What changed
ANCHOR evaluated 2,008 synthetic conversations across 27 personas, nine interaction schedules, three memory settings, and four models. Mean trajectory accuracy was 44.4%, user-state recall stayed near four-choice chance, and no context or memory setup consistently preserved continuity.
Why it matters
An assistant can sound coherent in each reply while gradually losing role boundaries, values, and consequential facts about the user.
What to do
Test assistants with sealed identity probes, counterfactual memory questions, and scheduled drift checks, while keeping critical user facts in validated state instead of conversational summaries alone.
Caveat
The study is synthetic and covers four models. Persona stability is evaluator-sensitive, and these scores do not directly measure user well-being or real-world outcomes.