Common Model of Cognition (CMC) Overview

• The Common Model of Cognition is a unified, modular architecture that integrates perception, memory, executive control, and motor actions to mimic human intelligence using both symbolic and neural approaches.
• It combines parallel perceptual processing with a serial production system, ensuring coordinated decision-making and adaptive responses across its distinct modules.
• Extensions of the CMC incorporate metacognition, emotion, and neuro-mimetic methods, enabling applications in interactive agents, advanced cognitive simulations, and neuro-symbolic integration.

The Common Model of Cognition (CMC) is a consensus, high-level computational architecture that encapsulates key principles and mechanisms underlying human-like cognition. It serves as a canonical blueprint for both cognitive modeling and the design of intelligent agents, integrating insights from cognitive science, artificial intelligence, and neuroscience. The CMC features a small set of functionally distinct, interacting modules, each implemented as an information-processing subsystem with defined interface protocols, learning mechanisms, and timing constraints. This model is foundational for unifying symbolic, connectionist, and neuro-symbolic cognitive architectures, and it underpins contemporary work on analogical reasoning, metacognition, neuro-mimetic systems, and the integration of generative neural networks.

1. Core Architecture and Formalization

The CMC is most commonly specified as comprising five core modules:

• Perception (P): processes external sensory input into internal encodings.
• Declarative Memory (DM): long-term storage and retrieval of factual and episodic information.
• Procedural Memory (PM): a production rule system acting as a central executive.
• Working Memory (WM): a centrally shared set of buffers for current percepts, goals, and intermediate inferences.
• Motor/Action (M): translates cognitive outputs into overt responses.
• (Optionally) Emotion (E): modulates the cognitive cycle via affective appraisals (Rosenbloom et al., 19 Dec 2024).

At the heart of the CMC is working memory, implemented as a set of named buffers $B = \{b_1, \ldots, b_k\}$, each holding a single symbolic chunk at a time. Chunks are finite feature–value structures $$c = \{\text{slot}_1 : v_1, \text{slot}_2 : v_2, \ldots\}$$.

Procedural memory realizes cognition via productions, each evaluated every $50\ \text{ms}$ in serial. Productions are rules of the form:

$$\text{if} \quad \forall i: b_i \text{ matches } P_i \quad \text{then (actions)}$$

Formally, the cognitive cycle is:

• At cycle $t$:
  • $$M_t \leftarrow \{p \mid \forall i: b_i(t) \approx P_i(p)\}$$
  • $p^* \leftarrow \arg\max_{p \in M_t} U(p)$
  • $\text{execute}(p^*)$, producing WM/memory/environment changes
  • $t \leftarrow t + 1$

Here, $U(p)$ is a utility function, often adapted online via temporal-difference (TD) learning. This sequencing imposes a serial bottleneck on central decision-making, while individual modules may process information in parallel (West et al., 25 Jan 2024).

2. Information Processing Cycle and Data Flow

The canonical CMC cycle orchestrates cognition as follows (Rosenbloom et al., 13 Jun 2025):

1. Perceptual modules (P) process sensory input and update their dedicated WM buffers.
2. DM receives cues from WM, returning retrieved facts or episodic traces into dedicated buffers.
3. PM monitors the complete set of WM buffers, matching condition patterns to select a single production (operator) to apply.
4. The selected production may request additional DM retrieval, update WM, or send output to the motor module.
5. The motor module issues commands to the environment.
6. Learning mechanisms operate continuously:
   • PM utilities are updated via TD reinforcement learning.
   • DM updates retrieval/activation metadata.
   • Behavioral outcomes feed into consolidation and memory updating.

This architecture ensures tight integration of parallel module processing with serial, globally coordinated executive control.

3. Module Specialization, Buffers, and Production Systems

Chunks, Buffers, and Productions

• Chunks are finite collections of role–filler pairs, supporting both symbolic and hybrid representations.
• Each buffer $b_i$ holds at most one chunk at $t$; production rule conditions specify joint patterns over all buffers.
• Productions fire at a fixed cycle (e.g., $50\ \text{ms}$), matching buffer conditions and issuing effects on WM, DM, or motor outputs.
• Central bottleneck: only one production fires per cycle; peripheral modules may implement parallel, local computation.

Serial and Parallelism

• Serial selection in PM guarantees interpretability and conflicts resolution.
• Buffers decouple parallel external input from the serial internal decision process, supporting robust event handling and concurrent context sensitivity (West et al., 25 Jan 2024, Conway-Smith et al., 2023).

4. Extensions: Shadow Production Systems, Middle Memory, and Generative Integration

West et al. (2024) introduce shadow production systems for peripheral modules, promoting a distributed but still hierarchically controlled architecture:

• Each peripheral module $X$ (P, M, E, DM) is endowed with a set of shadow productions $S^X$ acting in parallel and writing only to their dedicated buffers.
• The central PM then aggregates these intermediate results and selects a single central production for execution, preserving serial control (West et al., 25 Jan 2024).

A novel "Middle Memory" (MM) layer is introduced as an interface between symbolic WM buffers and large generative neural networks $G = \{G^1, ..., G^n\}$.

• On each cycle $t$, query vectors $x_t = f_1(WM(t)) \oplus f_2(MM(t))$ are constructed and sent to $G^k$ to produce predictive outputs $v_t^{(k)}$.
• Outputs are tagged and stored in MM, with activations reflecting recency, frequency, or spreading.
• MM and WM together drive both shadow and central production systems, supporting the seamless integration of symbolic and neural generative computations.

This extension enables neuro-symbolic architectures with modular interaction between symbolic reasoning, generative prediction, and parallel feature extraction (West et al., 25 Jan 2024).

5. Learning, Adaptivity, and Analogical Concept Memory

Learning in the CMC proceeds via both architectural and metacognitive routes:

• Production utilities in PM are adjusted using reinforcement-learning–style delta rules.
• Procedural compilation compacts repeated multi-step procedures into single productions.
• DM adjustments include incremental updates to base-level activation and associative strengths.

Analogical Concept Memory (ACM), as implemented in Soar (Mohan et al., 2020, Mohan et al., 2022), augments declarative learning by integrating analogical generalization engines (SME, SAGE):

• ACM supports relational, few-shot inductive concept formation via analogical similarity metrics and incremental generalization.
• It interleaves with episodic/semantic memory, providing a unified declarative substrate for recognition, planning, and language-grounded learning.

Agents using ACM in the AILEEN system displayed rapid few-shot concept acquisition, relational reasoning, and robust action induction with minimal guided examples, supporting both interactive task learning and embodied language processing (Mohan et al., 2020).

6. Metacognition, Emotion, and Dual-Process Spectrum

The CMC has been extended to explicitly account for metacognitive and affective processes (Rosenbloom et al., 19 Dec 2024, Laird et al., 9 Jun 2025):

• Metacognitive Assessment modules monitor internal module traffic and state, compute low-level appraisals (e.g., surprise, familiarity), and feed results into both WM and the Emotion module.
• The Emotion module computes an $n$-dimensional emotion vector $E$ from physiological, perceptual, and appraisal sources, modulating module communication, production selection, and retrieval thresholds.
• Pervasive bidirectional connections allow emotion to filter, gate, and amplify information flows, modulating both declarative and procedural computations.

System-1 and System-2 style cognition are emergent properties:

• Fast, automatic, heuristic (System-1) reasoning arises from direct production-firing in PM on current WM, with little DM retrieval.
• Deliberative, analytic, and metacognitive (System-2) reasoning involves multi-step manipulations of declarative chunks in WM and complex production sequencing.
• Importantly, there is no discrete “System-1” or “System-2” architectural component; the CMC provides a spectrum of processing modes arising from interaction patterns among core modules (Conway-Smith et al., 2023, Conway-Smith et al., 2023).

7. Neuro-mimetic and Predictive Processing Implementations

The CMC has been instantiated in neuro-mimetic frameworks combining Hebbian learning and variational free energy minimization:

• CoNGEN (COGnitive Neural GENerative system) organizes CMC modules as neurogenerative (NGC) circuits and associative memories (e.g., MINERVA 2, Hopfield).
• Perception and motor pathways are modeled as deep neural predictive coding networks, minimizing free-energy functionals; procedural and control circuits use Hebbian gating akin to basal ganglia mechanisms (Ororbia et al., 2023).
• Declarative and working memory are realized as high-dimensional vector or holographic reduced representations (HRR), supporting binding, unbinding, and superposition operations (Ororbia et al., 2021).
• All learning is effected through local, biologically plausible (two-factor) Hebb rules, unifying perception, control, and memory under a global predictive framework.

Empirical validation demonstrates robust performance on grid-world tasks, memory challenges, and continual learning benchmarks, with neuro-mimetic CMC models matching or exceeding deep RL baselines while maintaining interpretability and modularity (Ororbia et al., 2023, Ororbia et al., 2021).

8. Theoretical Integration and Application Domains

The CMC serves as a scaffold for integrating classical symbolic architectures (Soar, ACT-R, Sigma), hybrid neuro-symbolic systems, and contemporary theories of consciousness:

• Converging evidence links the CMC working memory module to the "global workspace" in consciousness frameworks, with PM enacting the executive spotlight (Global Neuronal Workspace Theory), symbolic structure as the substrate for integration (IIT), and predictive feedback as an explicit module-to-module channel (Predictive Processing) (Rosenbloom et al., 13 Jun 2025).
• Applications of the CMC include agent modeling, interactive task learning, behavior change interventions, and adaptive coaching, realized via modular system design that instantiates goal setting, context-aware reminders, and real-time feedback based on integrative judgments and experience (Mohan, 2019).

9. Limitations and Future Directions

Numerous extensions and open research problems remain:

• Full integration of neural generative modules with symbolic shadow production systems and metacognitive buffers requires further empirical work (West et al., 25 Jan 2024).
• Automated induction and adaptation of shadow productions, optimal encoding for mixed symbolic/vector representations, and efficient MM-GN interfacing are ongoing challenges.
• Empirical scaling to richer, compositional, and context-sensitive concept learning is required to extend ACM, MM, and neuro-mimetic systems.
• The CMC blueprint facilitates minimal and modular expansion (e.g., adding emotion or metacognition) without sacrificing architectural integrity, supporting community-driven consensus building on further extensions (Rosenbloom et al., 19 Dec 2024, Laird et al., 9 Jun 2025).

The Common Model of Cognition remains the prevailing architectural foundation for unified accounts of human-like intelligence, supporting incremental theoretical advances and practical cognitive systems across symbolic, neural, and hybrid paradigms.