The authors identify "self-locking" as a runtime failure mode in long-term persona agents, where generated lives collapse toward familiar environments and stale states due to model convergence and context gravity. To address this, they introduce AutoPersonas, a multi-timescale life-environment engine that separates environment-side Occurrences, accumulated Observations, and persona State within an OSO loop.
- A three-year compressed simulation exposed issues like occurrence-hardening gaps and recursive indecision in existing systems.
- An eight-model 40-day stress test showed mean rolling 5-day action-category repetition between 95.2% and 97.6%, with all models crossing 90% by day 11.
- Semantic re-keeping revealed 79.0%-88.0% macro-theme repetition across direct-loop runs.
- In a same-runtime 40-day A/B test, context-slice masking plus per-sample divergence targeting reduced macro-theme repetition from 61.8% to 36.3% and roughly doubled the cumulative theme count.
These results support the claim that separating controlled divergence from evidence-governed absorption can reduce persona-environment self-locking while preserving identity continuity.