Governed Hybrid Agency: A Critical Evidence Map of Cognitive Architectures for Foundation-Model Agents

Abstract

Background. Cognitive architectures specify the persistent memories, representations, learning processes, and control loops that allow an artificial agent to remain coherent across tasks and time. Foundation models provide broad generative competence, but do not by themselves ensure grounded action, reliable long-horizon memory, calibrated uncertainty, execution monitoring, or accountable learning.

Objectives. This critical integrative review examines how cognitive-architecture principles are being used to govern foundation-model agents. Taking the last comprehensive pre-foundation-model survey of cognitive architectures (Kotseruba & Tsotsos, first published online in 2018) as its historical baseline, it maps developments across twelve ability channels, identifies recurring integration patterns, and derives testable design and evaluation requirements.

Methods. We combined a historical-theoretical synthesis with a focused, PRISMA-informed evidence map of English-language records indexed in Scopus and the Web of Science Core Collection, searched in July 2026 over the 2018–2026 window. Queries required terms for cognitive architectures or architecture-level cognitive systems, foundation or generative models, and governance-relevant mechanisms. The searches identified 1,173 records; after deduplication, rule-based filtering, and LLMassisted title-and-abstract assessment, 119 works were retained and examined in full text. Because the 660 candidate records were not all independently human-rescreened, the corpus is interpreted as a reproducible focused evidence map rather than an exhaustive systematic review.

Results. All included works were published or available online in 2023–2026. Five recurring integration patterns were identified: foundation models used for knowledge or generative memory, language or multimodal interfaces, heuristic proposal or planning, rule or model bootstrapping, and bidirectional neuro-symbolic coupling. Across these patterns, a recurring design principle is to place foundation models inside persistent control loops that provide explicit state, provenanceaware memory, verification, metacognitive arbitration, execution monitoring, and recovery. Evaluation remains uneven, particularly for calibration, retention, transfer, long-horizon robustness, disturbance recovery, and reproducibility.

Conclusions. We propose governed hybrid agency, an Architectural Governance Stack, typed interface contracts, and a risk-based evaluation scorecard. These are integrative, falsifiable comparison targets rather than a validated universal architecture, and their value should be tested through matched ablations, intervention-sensitive evaluation, and cross-domain replication. A worked example instantiates the framework as a modular requirements-engineering reasoning architecture.