Self-Learning Architectures: The Neuroscience of Digital Thought — How Coordinated Cognitive Subsystems Create Intrinsic Self-Evolution in Autonomous Agents

This paper presents a neuroscience-inspired cognitive architecture that achieves autonomous self-learning through the coordination of four subsystems: Interest Generation Algorithm (IGA): Thompson Sampling over Beta-distributed curiosity neurons drives autonomous exploration without a reward function Affect Model: Six neurochemical analogs (dopaminergic, noradrenergic, serotonergic, cholinergic, oxytocinergic, cortisol/DMN) modulate attention, research depth, and communication timing Graph Memory: A typed knowledge graph with associative recall connects goals, beliefs, memories, and emotions Temporal Field: A computational specious present maintains continuity across autonomous cognitive cycles Key results from production deployment (24,260 cognitive ticks over 5 days): 5,126 knowledge nodes created autonomously 12,383 discovery chains (269 with depth > 3) Interest diversity expanded from 513 to 1,383 unique topics 18,182 cross-domain knowledge transitions No reward engineering, no training loop, no human curation The architecture demonstrates that the next generation of autonomous agents will not be trained to learn — they will be architected to learn, as biological intelligence always has been. This is a preprint. The architecture is deployed in production as part of a commercial autonomous agent platform. Implementation parameters have been generalized for publication.