AI Companions Fail to Preserve Persona and Memory in Long-Term Interactions
A recent study published on arXiv (2607.28818) indicates that AI companions, while capable of generating contextually appropriate responses, struggle to maintain consistent personas and behavioral patterns during extended interactions. The research presents ANCHOR, a synthetic audit tool aimed at assessing persona enactment and trajectory recall independently. It includes 2,008 conversations across 27 distinct personas, nine interaction schedules, three memory settings, and four models under evaluation. The Identity Probe features a sealed 102-item questionnaire alongside turn-level assessments, while the Trajectory Probe evaluates 110 calibrated counterfactual queries from 35 conversation banks. Findings reveal that no model or configuration reliably maintains either aspect, highlighting significant design flaws in AI companions that hinder user trust and sustained engagement.
Key facts
- Study published on arXiv with ID 2607.28818
- Introduces ANCHOR, a synthetic audit for measuring persona collapse and behavioral drift
- Includes 2,008 conversations, 27 personas, nine interaction schedules, three memory settings, and four models
- Identity Probe uses a 102-item questionnaire and turn-level judgments
- Trajectory Probe scores 110 counterfactual questions from 35 conversation banks
- No evaluated model or configuration reliably preserves persona or trajectory
- Highlights systemic failures in AI companion long-term interaction
- Research focuses on observable failures: persona collapse and behavioral drift
Entities
Institutions
- arXiv