Activation Arrays as Working Memory in Semantic Reasoning Systems

Description Architectural Primitive and Cognitive Mechanism This paper introduces Activation Arrays, a core primitive of the Semantic Topology Reasoning Architecture (STRA). Activation Arrays function as explicit, structured working memory for semantic reasoning systems, enabling transparent, inspectable, and evolvable reasoning processes. Unlike transformer-based models—where intermediate reasoning states remain hidden inside dense activations—Activation Arrays externalize reasoning into discrete, interpretable structures that can be inspected, validated, and corrected. The framework defines arrays as dynamic workspaces where concepts from the semantic topology are activated, transformed, filtered, ordered, and synthesized through explicit operations. These operations—SPREAD, FILTER, FORK, MERGE, ORDER, ANCHOR, and CONFESS—form a learnable reasoning vocabulary that bridges symbolic manipulation with neural adaptability. Arrays evolve through Hebbian dynamics, selection pressure, and meta-observation, enabling the system to refine its reasoning patterns over time. This work is part of the STRA Core Series, a coordinated research program exploring structure‑centric reasoning. While the STRA manifest presents the full architecture, this paper focuses specifically on working memory, procedural reasoning, epistemic honesty, and the evolutionary dynamics that allow reasoning systems to grow more coherent through use. Abstract Activation Arrays provide STRA with explicit, inspectable working memory. Arrays store activated concepts from the semantic topology and support reasoning through a set of discrete transform operations. These transforms—SPREAD, FILTER, FORK, MERGE, ORDER, ANCHOR, and CONFESS—enable exploration, pruning, hypothesis branching, synthesis, causal ordering, validation, and uncertainty expression. Arrays evolve through Hebbian strengthening, selection pressure, and meta-observation, allowing the system to learn both reasoning procedures and knowledge structure over time. This paper formalizes Activation Arrays as a hybrid symbolic–neural mechanism that supports transparent, adaptive, and epistemically honest reasoning. Background This work builds on the STRA manifest and complements the other papers in the STRA Core Series: Semantic Topology Reasoning Architecture (STRA) — 10.5281/zenodo.18207533 Causal Signatures for Cross‑Domain Reasoning — 10.5281/zenodo.18207546 Selection Pressure & Natural Selection of Thought — 10.5281/zenodo.18207552 Transform Learning: Procedural Reasoning Without Programs — 10.5281/zenodo.18207558 The Semantic Abacus — 10.5281/zenodo.18207560 Methodological foundations are documented in: The Practice of Human‑AI Synthesis — 10.5281/zenodo.17763521 The Minimal Knowledge Paradox — 10.5281/zenodo.17931472 Design as Epistemological Pathway — 10.5281/zenodo.18067554 Together, these works define a transparent, design‑led approach to AI reasoning. Key Contributions Explicit Working Memory: Defines Activation Arrays as structured, inspectable reasoning workspaces Transform Operations: Introduces a learnable vocabulary of reasoning actions (SPREAD, FILTER, FORK, MERGE, ORDER, ANCHOR, CONFESS) Parallel Reasoning: Formalizes Meta Arrays for maintaining multiple hypotheses simultaneously Epistemic Honesty: Provides mechanisms for uncertainty, contradiction detection, and confusion nodes Evolutionary Learning: Integrates Hebbian dynamics, selection pressure, and meta-observation for continuous refinement Reasoning Provenance: Enables step-by-step tracing of reasoning failures and successes Hybrid Cognition: Bridges symbolic manipulation with neural learning through transform selection Research Impact This work contributes to cognitive architectures, explainable AI, and reasoning systems by: Providing a transparent alternative to hidden-state reasoning Introducing explicit, evolvable working memory for AI systems Enabling interpretable procedural reasoning without programs Offering a foundation for debugging, verification, and error provenance Demonstrating how reasoning can evolve through use, not retraining Supporting multi-perspective analysis and pluralistic reasoning Keywords activation arrays, working memory, semantic reasoning, transparent AI, reasoning provenance, transform operations, meta-reasoning, epistemic honesty, Hebbian learning, selection pressure, cognitive architectures, STRA, semantic topology, interpretable AI, procedural reasoning Access and Documentation ORCID: https://orcid.org/0009-0003-4876-9273 GitHub: https://github.com/Neuron-Soul-AI/Neuron-Soul-AI Academia.edu: https://independent.academia.edu/MarceloTeixeira214 LinkedIn: https://www.linkedin.com/in/marcelo-emanuel-paradela-teixeira-702082382/ Email: marcelo.soul.ai@gmail.com License: CC BY-NC 4.0 © Marcelo Emanuel Paradela Teixeira 2026