ATHENA: A Cognitive Architecture for Persistent Execution and Skill Formation in Autonomous Agents

We present ATHENA (Adaptive Tiered Hybrid Execution and Neuro-Symbolic Architecture), a cognitive architecture for autonomous agents that compiles neural deliberation into persistent executable knowledge. ATHENA integrates neural and symbolic components in a three-tier hierarchy: Tier 3 (“Will”): a deliberative neural planner for novel problem solving; Tier 2 (“Habits”), an associative episodic memory that caches recurring solution traces; and Tier 1 (“Skills”), a procedural store of executable symbolic skills. Motivated by dual-process cognition and multi-system memory theories, ATHENA uses a neural planner (e.g. a Hierarchical Reasoning Model or LLMs) to solve novel problems while recording full execution traces for each solution; repeated regularities in these traces are embedded, clustered, and collapsed into symbolic skills or cached habits. Over time, computationally expensive deliberation is amortized into fast, deterministic execution, such that each successful inference permanently reduces future reasoning cost. We detail ATHENA’s architecture and learning mechanisms, including trace embedding, execution clustering, collapse criteria, and skill induction, and compare it to classical cognitive architectures (Soar, ACT-R, LIDA, CLARION, etc.), highlighting its explicit neuro-symbolic execution pipeline, decision traceability, and precedent-constrained skill learning. Finally, we discuss implications for bounded rationality, heuristic formation, and explainability, and demonstrate applicability to ARC-AGI tasks, robotics, and structured reasoning domains, showing how ATHENA enables efficient, transparent, and self-improving intelligence.