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

Abstract

Large Language Model (LLM) agents demonstrate remarkable capabilities but struggle with self-evolution the ability to autonomously improve their own cognitive architecture through experience. We present STOS (Seek Truth OS), a biomimetic cognitive architecture that enables self-evolving LLM agents through three core innovations: (1) a three-layer belief system with hierarchical drift mechanisms that allow beliefs to strengthen or weaken based on experience; (2) a metacognitive audit system that monitors belief accuracy and detects anomalies; and (3) an excretory system that removes outdated or low-confidence knowledge, preventing cognitive pollution. We validate STOS through four experiments: belief drift correction (verifying hierarchical relationships), metacognitive audit detection rates (precision/recall/F1), excretory system cleanup rates, and ablation studies quantifying each modules contribution. Results show that STOS achieves 85% belief drift accuracy, 78% metacognitive precision, and 73% excretory cleanup rate, with the three-layer belief system contributing most significantly to overall performance. STOS represents a step toward truly autonomous cognitive systems that can evolve without human intervention.