Researchers have identified a critical failure mode in long-term AI persona agents called "self-locking," where simulated characters gradually collapse into repetitive behaviors despite appearing locally plausible.

The phenomenon occurs when AI agents converge toward high-probability behavioral patterns, leading to familiar environments, weak relationships, and stagnant decision-making over time. Author Mengchen Li traced this to model-level convergence and system-level "context gravity" from accumulated state and memory summaries.

AutoPersonas addresses this through a multi-timescale architecture that separates environment-side occurrences, accumulated observations, and persona state. The system's OSO loop allows for divergent future scenarios while requiring evidence-based validation before making state changes.

Testing Reveals Widespread Repetition

A comprehensive 40-day stress test across eight different AI models generated 1,600 events and revealed concerning patterns. Mean rolling 5-day action-category repetition ranged from 95.2% to 97.6%, with all models crossing the 90% threshold by day 11.

Semantic analysis found macro-theme repetition between 79.0% and 88.0% across direct-loop runs. The research team conducted a three-year compressed simulation that exposed multiple failure modes including environment watermark shells and recursive indecision patterns.

In controlled A/B testing, context-slice masking combined with per-sample divergence targeting reduced macro-theme repetition from 61.8% to 36.3%. The intervention roughly doubled cumulative theme count while maintaining identity continuity.

A fictional-world experiment using a juvenile-goblin character successfully reproduced the anti-fixation regime without real-world constraints, suggesting the approach works across different simulation contexts.

The 52-page paper includes 13 figures and tables, with supplementary evaluation artifacts made publicly available. The research addresses a fundamental challenge in developing AI agents capable of sustained, realistic long-term behavior simulation.

The findings have implications for AI companion development, virtual world simulation, and any application requiring persistent AI characters that must evolve naturally over extended periods while maintaining core identity traits.