STOS: A Biomimetic Cognitive Architecture for Self-Evolving LLM Agents

Abstract

Large Language Model (LLM) agents demonstrate remarkable capabilities but lack mechanisms for autonomous self-evolution—the ability to improve their own cognitive structures through experience. We present STOS (Seek Truth OS), a biomimetic cognitive architecture that integrates hierarchical belief drift, metacognitive auditing, and knowledge excretion into a unified framework for self-evolving LLM agents. The belief system organizes knowledge at three hierarchical levels (meta, core, application) with level-dependent learning rates that enforce differential plasticity: meta-level beliefs drift slowly (5% over 50 rounds), core beliefs at moderate rates (25%), and application-level beliefs adapt most rapidly (50%). A metacognitive audit module monitors belief accuracy and detects anomalies, achieving perfect detection on a controlled evaluation (precision = recall = F1 = 1.0 on 30 entries with 10 anomalies). An excretory subsystem removes stale, contradictory, and zombie knowledge items, achieving 100% cleanup with 0% false-positive collateral damage across 14 target items. Ablation experiments confirm layer orthogonality. This paper presents Part I of a multi-part research program.