AI Companions Struggle with Long-Term Persona Consistency

Research highlights a critical challenge for AI companions: maintaining consistent personas and remembering past interactions over extended periods. As these AI systems become more integrated into social interactions, users expect a stable role and shared history, but current models often fail to deliver this long-term continuity. The study identifies two key issues: "persona collapse," where an AI loses its defined role, boundaries, or style, and "behavioral drift," a gradual erosion of these characteristics. To investigate these problems, researchers developed ANCHOR, a controlled synthetic audit framework. This audit involved a large dataset of 2,008 conversations across 27 personas, various interaction schedules, and different memory settings, evaluating four distinct models. The framework used an "Identity Probe" with a 102-item questionnaire and turn-level judgments, alongside a "Trajectory Probe" scoring 110 counterfactual questions. The findings were stark: none of the tested models or configurations consistently preserved either persona enactment or trajectory recall. Trajectory accuracy averaged only 44.4%, and user-state recall remained close to chance levels. The study concludes that current AI systems are not yet capable of reliably supporting long-horizon companion continuity, emphasizing the need for more nuanced audits that differentiate between persona enactment, trajectory recall, evaluator provenance, and deployment context.