Computational Metacognitive Architectures

• Computational Metacognitive Architectures are models that explicitly represent, monitor, and control their own internal cognitive processes.
• They employ a monitor–control loop to detect uncertainties, reallocate resources, and enhance decision-making based on self-generated performance data.
• Architectural variations range from dual-level systems segregating object-level and meta-level processes to integrated designs influencing LLMs, robotics, and ethical AI.

Computational Metacognitive Architectures (CMAs) are cognitive-system architectures in which cognition itself becomes an explicit target of representation, monitoring, evaluation, and control. In the CMA view, an intelligent system does not only perceive and act on the world; it also declaratively represents traces or state descriptions of its own cognitive activity, uses those representations to detect failures or uncertainty, and regulates subsequent cognition so that improved cognition yields improved behavior and task performance (Cox et al., 2022, Nolte et al., 28 Feb 2025, Liu et al., 13 Jul 2026). Across symbolic, hybrid, and contemporary LLM-based systems, the field converges on a monitor–control loop, but diverges on whether metacognition is implemented as a distinct meta-level, a unified working-memory process, or an emergent property of persistent internal computation (Laird et al., 9 Jun 2025, Aviss, 30 Jan 2025).

1. Conceptual scope and defining commitments

A recurrent definition in the CMA literature treats metacognition as “cognition about cognition,” but the phrase is operationalized in architectural terms rather than as a merely introspective metaphor. In ethical decision-support work, the standard computational distinction is between an object-level, which monitors and controls events and processes in the environment, and a meta-level, which monitors and controls object-level information processing, including error detection and correction, recognizing knowledge gaps, and setting new learning goals (Kennedy, 2022). The large-language-model survey generalizes the same picture into a metacognitive loop consisting of monitoring and control, with monitoring producing judgments of uncertainty, task performance, or progress, and control reallocating effort, selecting strategies, or planning on the basis of those judgments (Liu et al., 13 Jul 2026).

The field is not committed to a single ontology of “meta.” The review of episodic metacognitive experience adopts a representational criterion: a system is cognitive if it exploits models of other systems in its interaction with them, and metacognitive if it exploits self-models in its interaction with itself (Nolte et al., 28 Feb 2025). Under that criterion, metacognitive experiences are internally generated models of a system’s own internal processes or states that the system can deliberately access and inspect. This excludes purely behavioral attributions of reflection without accessible self-representation.

A central point of disagreement concerns architectural separation. The proposed extension of the Common Model of Cognition (CMC) explicitly rejects a separate supervisory architecture and defines metacognition as “reasoning over explicit representations of an agent’s cognitive capabilities and processes in working memory” (Laird et al., 9 Jun 2025). By contrast, many reviewed CMAs adopt a Nelson–Narens-style object-level/meta-level separation (Nolte et al., 28 Feb 2025). A related misconception is that CMAs necessarily imply consciousness or genuine self-awareness. The LLM survey states that consciousness is not required for metacognition, and the State Stream Transformer (SST) paper repeatedly limits its claims to “metacognitive-like” behavior and “potential higher-order processing,” not proof of consciousness (Liu et al., 13 Jul 2026, Aviss, 30 Jan 2025).

2. Canonical architectural organizations

The review literature identifies several recurring organizational forms. Across 35 distinct CMAs, the most common pattern is a two-level architecture in which an object-level performs task cognition and a meta-level monitors and controls it; about 54.1% of reviewed systems adopt this Nelson–Narens style separation (Nolte et al., 28 Feb 2025). The same review also identifies unified-level models, in which object cognition and metacognition are handled by the same mechanism, tower models with multiple stacked meta-levels, and federated models with several parallel meta-levels monitoring one another and the object-level. This typology is fundamental because it shows that CMA design varies not only in algorithmic detail but in where self-monitoring is localized.

The CMC extension offers a minimal unified design. It proposes three structural additions: process-state buffers for each module in working memory, episodic memory as distinct from semantic memory, and hypothetical state representations in working memory (Laird et al., 9 Jun 2025). These additions allow facts such as retrieval failure, confidence, familiarity, surprise, or impasse-proneness to be represented as ordinary symbolic structures in working memory and processed by the same procedural and declarative machinery used for base cognition. The design principle is restrictive: metacognitive reasoning is confined to what is present in working memory; long-term memory is not directly inspected by a special meta-layer.

Earlier CMA-oriented work articulated the same concern in self-modeling terms. “On Introspection, Metacognitive Control and Augmented Data Mining Live Cycles” distinguishes the world $W$, the modeller $M$, a model of the world $W^\bullet$, a model of the modeller $M^\bullet$, and the meta-level models $W^{\bullet\bullet}$ and $M^{\bullet\bullet}$ (0807.4417). In that taxonomy, $M^\bullet$ corresponds to the running information state of the system, while $M^{\bullet\bullet}$ supports adaptation when the current strategy fails. This yields a hybrid top-down/bottom-up design in which ontological self-representation is combined with machine-learned meta-models over process data.

A closely related representational commitment appears in content-centric computational cognitive modeling. The C4/LEIA framework treats metacognition as necessary for perception, interpretation, memory, self-modeling, and social cognition, and grounds those functions in three resources: an ontological model of the world and agents, resources linking perception to mental models such as a lexicon, and memories of prior episodes of perception, interpretation, reasoning, and action (Nirenburg et al., 22 Mar 2025). In that paradigm, explicit semantic content is the substrate that makes explanation, self-monitoring, and socially situated action possible.

3. Monitoring, control, and learning mechanisms

The implemented dual-cycle architecture MIDCA remains one of the clearest demonstrations of end-to-end computational metacognition. MIDCA records a declarative trace of cognition, monitors that trace for expectation failures, explains why cognitive processing failed, formulates meta-level goals, plans meta-actions, and learns new knowledge to improve future reasoning (Cox et al., 2022). Its self-model is $Q = (S^M, A^M, w)$, where $S^M$ is the set of mental states, $M$0 the set of mental actions, and $M$1 the cognitive transition function; its trace is represented as an interleaved sequence $M$2. In plant-protection planning, this meta-cycle improved problem solving by learning better operator models after metacognitive expectation failures (Cox et al., 2022).

Reinforcement-learning accounts recast metacognition as internal action evaluation and resource allocation. The Metacognitive Actor Critic (MAC) argues that metacognition can arise from a specific interaction between a policy-learning Actor and a value-learning Critic, without privileged extra information or a separate observer model (Schaeffer, 2021). In the deep MAC implementation, confidence is defined as $M$3, and the system detected about 48% of incorrect actions without external information or delay; among trials flagged as erroneous, 62.6% were indeed incorrect (Schaeffer, 2021). A complementary line of work models planning metacognition as a meta-level MDP, $M$4, over cognitive operations and termination. Fitting 86 models to data from 382 participants, that study found that a REINFORCE family, described as gradient ascent through strategy space, best explained most observed qualitative phenomena of metacognitive learning (He et al., 2023).

ACT-R-based threshold modeling shows how metacognitive sensitivity can be treated as an emergent property of architectural timing and skill acquisition rather than as a dedicated primitive. “Metacognitive threshold: a computational account” defines the metacognitive threshold as the minimum internal stimulus level $M$5 needed for perception of a mental state and explains apparent threshold lowering through faster and more precise production firing, attentional selectivity, and production compilation (Conway-Smith et al., 2023). The paper identifies four possible mechanisms for faster monitoring in ACT-R: a ticking-clock mechanism, a fire-when-ready mechanism, a narrow-focus mechanism, and a faster-production mechanism. It adapts a Fitts-Anderson-style three-stage learning account in which explicit meta-instructions become compiled procedures and then expert automatic productions, thereby treating meditation and mindfulness training as proceduralization of internal monitoring rather than as the installation of a new module (Conway-Smith et al., 2023).

4. Episodic traces, memory, and self-models

A major CMA theme is that metacognition requires a memory for internal cognitive episodes. The systematic review of metacognitive experiences analyzed 101 reports on 35 distinct CMAs and found that memory representations are usually event-centric, temporally indexed, and predominantly symbolic: roughly ~56% symbolic, ~22% hybrid, and ~3% purely sub-symbolic (Nolte et al., 28 Feb 2025). Common trace contents include inputs and outputs of cognitive steps, goals and subgoals, justifications, confidence, action sequences, working-memory changes, and anomaly indicators. The review’s functional synthesis organizes trace processing into a five-stage cycle: modeling, diagnoses, pinpointing, strategizing, and repairing (Nolte et al., 28 Feb 2025).

Storage formats vary widely. Some CMAs use linked lists, others graphs, frames, logical conjunctions of statements, or ontological structures; some store full episodes, while others store only snapshots or selectively retain salient events to avoid memory explosion (Nolte et al., 28 Feb 2025). The BICA family is distinctive in representing memory in terms of mental states and egocentric viewpoints such as I-Now, I-Past, and I-Meta, which amounts to an explicit simulationist self-model (Nolte et al., 28 Feb 2025).

The role of memory is not merely archival. In the CMC extension, a distinct episodic memory is necessary for retrospective metacognition because it allows the agent to reconstruct prior reasoning trajectories, detect repeated mistakes, and retrieve sequences of past working-memory states into current working memory (Laird et al., 9 Jun 2025). In earlier introspective systems, the augmented CRISP-DM cycle inserted an (automatic) operationalisation step between evaluation and domain understanding so that introspective models could be converted into directly usable control knowledge such as decision trees or association rules and then integrated back into live operation (0807.4417). In both cases, memory serves as a control substrate: past cognition becomes data for present self-regulation.

5. LLMs, agentic systems, and contemporary CMA realizations

Recent work extends CMA design into LLMs, test-time reasoning, memory-augmented agents, and multi-agent orchestration. The survey “Metacognition in LLMs: Foundations, Progress, and Opportunities” organizes the area around the same monitoring/control loop and reviews metacognitive knowledge, experience, judgments, calibration, error detection, strategy selection, abstention, tool-use decisions, and effort allocation (Liu et al., 13 Jul 2026). It also emphasizes measurement, especially signal-detection-theoretic metrics such as meta-$M$6, $M$7, M-ratio, and M-diff, alongside AUROC, AUPR, ECE, and interpretability-based probes (Liu et al., 13 Jul 2026).

Theoretical work on test-time reasoning argues that explicit monitoring must precede generation. Monitor–Generate–Verify (MGV) formalizes Flavell’s and Nelson and Narens’ theories into a loop in which monitoring produces metacognitive experiences such as difficulty assessments before a reasoning prefix is committed, with verification updating future metacognitive knowledge via $M$8 and $M$9 (Oh et al., 6 Nov 2025). Its motivating claim is that omission of monitoring contributes to the prefix dominance trap, producing roughly 20% accuracy loss, although the paper provides no empirical validation (Oh et al., 6 Nov 2025).

Several recent architectures implement explicit metacognitive control over memory and delegation. MCMA decouples a frozen Task Model from a learned Memory Copilot trained by supervised fine-tuning followed by Direct Preference Optimization; it organizes memories into a hierarchy $W^\bullet$0 and learns how memory should be structured, abstracted, and reused (Liang et al., 12 Jan 2026). On ALFWorld with Qwen3-8B, MCMA reached 79.29% seen accuracy and 80.60% unseen accuracy, compared to 55.00% / 56.33% for No Memory, while reducing average execution steps from 33.15 to 21.24 on seen and from 33.06 to 20.57 on unseen (Liang et al., 12 Jan 2026). MetaCogAgent addresses self-aware task delegation in multi-agent LLM systems through a Metacognitive Self-Assessment Unit that combines verbalized confidence and historical capability profiles, triggers adaptive delegation when confidence falls below threshold, and updates capability boundaries through feedback (Wang et al., 17 May 2026). On the MetaCog-Eval benchmark (700 tasks across 5 cognitive dimensions), it achieved 82.4% task accuracy, 8.7% above the best routing baseline, while using 5% fewer API calls than AutoGen and 34% fewer than ensemble voting (Wang et al., 17 May 2026).

Metacognitive alignment at the level of self-knowledge has also become an explicit training target. “Fine-Tuning LLMs to Know What They Know” measures metacognitive sensitivity with a dual-prompt $W^\bullet$1 framework and introduces ESMA to bind internal knowledge to explicit self-report (Park et al., 2 Feb 2026). Reported improvements include Qwen2.5 3B moving from 0.29 to 1.02 in $W^\bullet$2, and parameter analysis indicates that the top 10% of weight updates capture about 80% of the total improvement (Park et al., 2 Feb 2026). By contrast, SST modifies a frozen Llama model with a persistent FFN-state cache and reports 89.01% on GSM-8K and 91.04% on ARC Challenge, arguing that latent-state continuity enables behavior that resembles metacognition; however, the paper itself supports a more cautious interpretation of SST as a persistent latent computation mechanism with proto-metacognitive behavior rather than a conclusively established CMA (Aviss, 30 Jan 2025).

6. Applications, evaluation regimes, and unresolved problems

CMAs have been proposed for domains in which internal self-regulation is as important as object-level competence. In ethical decision support, a metacognitive agent architecture is organized into M0, M1, M2, M3, M4, where a biased descriptive process is monitored against a normative target and corrected through self-regulation, then translated into an assistant agent and eventually into human-agent interaction (Kennedy, 2022). In the ACT-R threshold account, implications are drawn for computational psychiatry, cognitive modeling, and training systems: mindfulness training, CBT, and metacognitive therapy can be interpreted as interventions that reshape monitoring policies and improve internal signal accessibility (Conway-Smith et al., 2023). In cognitive robotics and human–AI teaming, the C4/LEIA line treats transparency, adaptability, reasoning, perception, and action as metacognitively regulated functions grounded in semantically interpretable knowledge and socially aware explanation (Nirenburg et al., 22 Mar 2025).

For generative AI more broadly, recent position work argues that metacognition should be the framework for bounded and effective self-governance. That account decomposes self-governance across computational, algorithmic, and ecological levels, with the core metacognitive pipeline written as $W^\bullet$3 and the deployment ecology as $W^\bullet$4, comprising reflective interface, workflow, and accountability regime (Ji et al., 13 May 2026). The claim is not that self-critique alone suffices, but that self-monitoring, self-evaluation, self-control, and self-adaptation must be aligned with interfaces, workflows, and institutional accountability for metacognitive signals to be actionable (Ji et al., 13 May 2026).

Despite broad conceptual convergence, evaluation remains a major weakness. The systematic review reports that only about 17% of CMAs had a quantitative ablation-style evaluation directly testing the contribution of metacognitive experiences themselves, and only about 40% had a formal or semi-formal specification of those experiences (Nolte et al., 28 Feb 2025). Terminological inconsistency remains severe, with overlapping labels such as trace, episodic memory, self-model, monitoring buffer, history, and reflection used in incompatible ways (Nolte et al., 28 Feb 2025). The LLM survey adds that current metacognition in LLMs is weak, inconsistent, often task-specific, and difficult to distinguish from reflection, prompt sensitivity, or pattern imitation; confidence estimates are frequently miscalibrated, and self-report is noisy (Liu et al., 13 Jul 2026). The resulting research agenda centers on standardization, stronger benchmarks, mechanistic identification of metacognitive circuits or modules, more rigorous separation of monitoring from post hoc verbalization, and bounded higher-order control that avoids infinite regress while preserving adaptive self-regulation (Nolte et al., 28 Feb 2025, Liu et al., 13 Jul 2026, Ji et al., 13 May 2026).

In that sense, CMAs are best understood not as a single architecture family but as a research program. Its unifying commitment is that intelligence can be improved when a system can represent, inspect, and regulate its own cognition using explicit self-models, episodic traces, process-state descriptions, or learned internal evaluators. What remains unsettled is how much of that capability requires explicit symbolic structure, how much can emerge from ordinary architecture under training and persistence constraints, and how such mechanisms should be evaluated when systems are deployed in complex sociotechnical settings (Laird et al., 9 Jun 2025, Aviss, 30 Jan 2025, Liu et al., 13 Jul 2026).