AutoPersonas: A Multi-Timescale Loop Engine for Open-Ended Persona Evolution

Jul 16, 2026

Upload by Aurelia

Long-term persona agents must stay identifiable while adapting to new events, relationships, evidence, and social conditions. The work identifies self-locking as a runtime failure mode in continuing persona-life loops, where plausible local events mask a collapse toward familiar environments, weak ties, stalled decisions, and stale life stages. AutoPersonas introduces a multi-timescale engine separating Occurrences, Observations, and persona State, using an OSO loop with evidence-governed absorption to reduce self-locking while preserving identity continuity.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages51

Reading Time2 h 9 min

Abstract

Introduction

arXiv :2607 .08252v 1 [ cs .AI] 9 Jul 2026 AutoPersonas: A Multi-Timescale Loop Engine for Open-Ended Persona Evolution Mengchen Li Latrix [limengchen@latrix.ai](limengchen@latrix.ai) July 8, 2026 Abstract Long-term persona agents must remain identifiable while adapting to new events, relationships, evidence, and social conditions. We identify self-locking as a runtime failure mode in continuing persona-life loops. In this failure, locally plausible events keep appearing while the generated life collapses toward familiar environments, weak relationships, suspended decisions, and stale life stages. We trace this failure to two coupled pressures: model-level convergence toward high-probability behavioral channels and system-level context gravity from State, memory, history, and environment summaries. We introduce AutoPersonas, a multi-timescale life-environment engine for bounded persona-level recursive self-evolution. It separates environment-side Occurrences, accumulated Observations, and persona State. Its OSO loop lets divergent future-facing material enter the system while requiring evidence-governed absorption before State or reachability changes. We evaluate the architecture through diagnostic audits rather than benchmark superiority. A three-year compressed simulation exposed environment watermark shells, occurrence-hardening gaps, slowchange accumulation failures, recursive indecision, and weak relationship persistence. An eight-model 40-day action-only stress test generated 1,600 events. Mean rolling 5-day action-category repetition was 95 .2%-97 .6%; all included models crossed 80% by day 9 and 90% by day 11 . A semantic rekeeping of the same direct-loop outputs found 79.0%-88.0% macro-theme repetition across all eight models. In a same-runtime 40-day A/B, context-slice masking plus per-sample divergence targeting reduced macro-theme repetition from 61 .8% to 36 .3% in the masked lane. The intervention also roughly doubled cumulative theme count from 55 to 102 . A separate juvenile-goblin fictional-world run reproduced the anti-fixation regime with 42.8% blended repetition and no hard real-world intrusions. These results support a bounded systems claim: separating controlled divergence from evidence-governed absorption can reduce persona-environment self-locking while preserving identity continuity. 1 Introduction Large language model (LLM) agents increasingly persist across interactions. A long-term persona agent should do more than remember past conversations or maintain a stable role. It should appear to have a continuing life: entering new situations, responding to external changes, forming relationships, revising concerns, and moving through life stages. A persona that never changes maybe coherent, but it is not long-term. This setting differs from a common long-horizon-agent regime. Many long-horizon agents are organized as task-oriented systems: they receive objectives, interact with tools or users, obtain feedback, and improve by correcting errors against an external success criterion. AutoPersonas targets a different object. It is not an extended task episode, but a self-running persona entity that should continue between human instructions, repair its own understanding of the outside world, and revise its life-environment through its own loop. This is not a superiority claim; it is a difference in target and architecture. In task-oriented agents, external goals and evaluations can pull the system back toward a defined outcome. In persona open-evolution, no complete benchmark or scalar reward specifies what a human-like life should become, so the loop must create, absorb, and audit its own sources of growth. The target condition is therefore reward-free open-evolution under context gravity: the system must keep moving without an external objective function while its own memory, State, and environment summaries continually pull future generation back toward prior attractors. Persona life

FAQs

What is self-locking in continuing persona-life loops?

Self-locking is a runtime failure where locally plausible events keep generating while the persona’s overall life collapses toward familiar environments, weak relationships, suspended decisions, and stale life stages.

What pressures cause self-locking in the system?

The paper attributes it to two coupled pressures: model-level convergence toward high-probability behavioral channels and system-level context gravity from State, memory, history, and environment summaries.

How does AutoPersonas reduce self-locking while keeping identity continuity?

AutoPersonas separates environment-side Occurrences, accumulated Observations, and persona State, and uses an OSO loop that allows divergent future-facing material to enter only after evidence-governed absorption before State or reachability changes.

Jul 16, 2026

Upload by Aurelia