Cognitive Mirrors: The RCS Triad for Retrieval,Control, and Stewardship in Neuro-Inspired LLMSystems
As large language model systems evolve from standalone text generators into retrieval-enabled, tool-using, and multi-session agents, researchers increasingly turn to cognitiveneuroscience for interpretive guidance. This paper argues that such comparisons areuseful only when they are bounded, functionally specified, and empirically falsifiable. Weanalyze retrieval-augmented generation as a separation between parametric knowledgeand external evidence access, and we analyze agent runtimes as control layers responsiblefor planning, routing, tool use, and context management. We distinguish these runtimefunctions from Model Context Protocol, which is an interoperability standard for exposingresources, prompts, and tools rather than a theory of memory or executive control. On thisbasis, we propose the RCS Triad Retrieval, Control, and Stewardship as a compactframework for describing modern LLM systems. Retrieval concerns evidence access andmemory extension; Control concerns orchestration, tool selection, and active context;Stewardship concerns provenance, consent, auditability, risk boundaries, and humanoversight. The paper’s central claim is methodological rather than biological: neuroscienceshould not be used to assert one-to-one equivalence between brains and models, but togenerate disciplined architectural hypotheses. We position the RCS Triad against eightadjacent retrieval and agent frameworks in a comparative table, specify its boundaryconditions through a Convergence-Divergence Matrix across eight cognitive domains,and outline a benchmarkable empirical agenda with four falsifiable predictions. We alsointroduce two original governance concepts architectural negligence and goaltransparency as frameworks for understanding the liability and rights questions thatarise when AI systems confabulate and persist goal-states.
