Cognate: A Controller-Based Cognitive Architecture for Experience-Guided AI Agents

Contemporary AI agent frameworks are largely built on iterative prompt loops, combining large language models (LLMs) with tool invocation in reactive pipelines. While effective for narrow tasks, these systems lack structured cognition, persistent reasoning state, and principled execution control. This paper introduces Cognate, a controller-based cognitive architecture developed within the Thoth system. Cognate decomposes intelligence into distinct subsystems—planning, execution, memory, and learning—connected through a continuous, state-driven reasoning loop. The architecture demonstrates empirical signals of automated, experience-driven adaptation and stable iterative reasoning.