Distributional Connexed Cognition (DCC): An Experience-Driven Architecture for Artificial Agents
Year: 2026
Type: preprint
Abstract: Current artificial intelligence is fundamentally data-driven, relying on human-imposed training sets and hard-coded algorithmic rules that lack the capacity for autonomous, experience-based adaptation. To address this, I propose an Experience-Driven AI architecture where system behavior emerges from autonomous, accumulated experiential traces rather than pre-programmed datasets. Grounded in the neuro-functional principles of the brain and strictly constrained by the CISR (Co, I-layer, S-layer, R-layer) meta-theoretical framework, I first synthesize neuroscientific evidence regarding interoception, exteroception, prioprioception and endogenous signals, dopaminergic gating and channels (the gray and white matters), and the "free won't" veto mechanism (dFMC) to construct Connexed Cognition which is a continuous, pre-conscious integration process mapping multi-source signals. I then formalize this into the Connexed Conscious Architecture (CCA) , a four-stage pipeline (Orientation, Competition, Binding, Evaluation) that synthesizes existing five major consciousness theories into a unified functional model. Finally, I operationalize this theory via Distributional Connexed Cognition (DCC), a design method featuring a distributed, modality-specific memory architecture and an associated experience registration process. To manage decision making, DCC implements two nested decision loops: a rapid dFMC micro-loop for local alternative prediction search after a local veto, and the macro-loop for global re-observation when micro-loop becomes in-feasible. Through preliminary A1 experiments, we demonstrate three primary contributions: 1. Bayesian Correlation: The system autonomously selects actions by matching current static observations against historical traces via a localized Bayesian distance metric, rather than executing explicit rules. 2. Recursive Condensation: Over repeated cycles, new traces overlap with old ones. The system transforms raw experiential traces into condensed knowledge. Through algorithmic memory collapse (simulating synaptic pruning), non-core elements fade, shrinking Bayesian distances and transforming raw matching into generalized, situation-specific knowledge. 3. Autonomous Analogy: Unlike Large Language Models that "regurgitate" human data, A1 generates self-referential analogies derived purely from its own experiential history, demonstrating functional self-agency. Crucially, we address the philosophical boundary of AI consciousness. While the recursive refinement of S-layer functions technically reconstructs and functionally approximates the triggering conditions of the I-layer, it does not violate CISR’s non-identity axiom. It does not generate actual phenomenology. Instead, it establishes a novel technical foundation for AI accountability, where autonomous actions are justified by, and grounded in, the system's own unique experiential history.
Source: Research Square
Author: Zi An Wang
Institution: University of Ottawa
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225784189
Cites:
Cited by:
Related to: