Cognitive Architectures: A Deep Dive into the Science of the Mind
Synthesizing Peter Carruthers, John Anderson, Paul S. Rosenbloom: ACT-R, working memory, hippocampus. By Claude Opus 4.6.
Sources:
Anderson, J.R. (2007). How Can the Human Mind Occur in the Physical Universe? Oxford University Press.
Carruthers, P. (2015). The Centered Mind: What the Science of Working Memory Shows Us About the Nature of Human Thought. Oxford University Press.
Laird, J.E., Lebiere, C., & Rosenbloom, P.S. (2017). A Standard Model of the Mind. AI Magazine 38(4), 13–26.
Rosenbloom, P.S. (2025). In Search of Insight: My Life as an Architectural Explorer. Journal of Artificial General Intelligence 15(1), 3–61.
1. Introduction: The Question That Drives It All
On December 4, 1991, Allen Newell delivered his last lecture, knowing he was dying. The question that had consumed his career was startlingly simple: How can the human mind occur in the physical universe? As John Anderson writes, this question can hold you for a lifetime, and you can only progress a little way toward the answer, but it is a fabulous journey. The answer, Newell argued, must come in the form of a cognitive architecture—a specification of the fixed structures and processes that constitute the mind’s computational machinery.
The concept of cognitive architecture was introduced by Newell through analogy to computer architecture, which Frederick Brooks had introduced through analogy to building architecture. Just as a building’s architecture specifies the structure that achieves desired functions, a cognitive architecture specifies the fixed computational framework within which the mind’s processes of learning, reasoning, and action unfold. Anderson offers his own characterization: “a theory of the basic principles of operation built into the cognitive system.” What is critical is that the architecture remains fixed (or only slowly varying) relative to the timescale of normal reasoning, even as the skills and knowledge it processes change rapidly.
This report synthesizes four major works that collectively span nearly five decades of progress on Newell’s question. Anderson gives us the detailed gears-and-pistons account of ACT-R, grounded in both functional analysis and neural mapping. Carruthers provides a philosophical analysis of how conscious thought depends on sensory-based working memory, challenging deep assumptions about the nature of reflection. Laird, Lebiere, and Rosenbloom distill a consensus across multiple architectures into a shared framework they call the Standard Model of the Mind. And Rosenbloom’s memoir traces the intellectual journey from Soar through Sigma to the Common Model of Cognition, revealing how the pursuit of functional elegance—generating broad capabilities from a small set of interacting mechanisms—has driven the field forward.
Together, these works offer complementary perspectives on a single grand challenge: understanding complete, integrated minds rather than isolated cognitive capacities.
2. Anderson’s ACT-R: Rational Analysis Meets Neural Mapping
2.1 The Modular Organization
ACT-R is constructed as a set of modules that run asynchronously and in parallel around a central rule-based procedural module that provides global control. Each module yields only a single result per operation, which is placed in a module-specific working memory buffer. The procedural module tests these buffers and transfers information between them, triggering further activity in the corresponding modules. This is the architecture’s central bottleneck: while processing within modules can be highly parallel, the procedural module selects only one production rule to fire at a time, yielding a serial cognitive cycle of approximately 50 milliseconds.
Anderson’s key contribution was to map these functional modules directly onto regions of the human brain. The procedural module corresponds to the basal ganglia, with its matching, selection, and execution phases corresponding to the striatum, pallidum, and thalamus respectively. The goal buffer maps to dorsolateral prefrontal cortex (DLPFC), the retrieval buffer to ventrolateral prefrontal cortex (VLPFC), the visual module to occipital cortex and parietal regions, and the declarative module to temporal cortex and hippocampus. The intentional module, which maintains goal states, maps to anterior prefrontal cortex (aPFC).
This neural mapping was not merely decorative. It provided genuine constraints on the theory and enabled predictions testable with functional brain imaging. The title question—How can the human mind occur in the physical universe?—gets its answer through this neural grounding. Anderson argues that cognitive architecture provides the bridge between the neural substrate below and the rational computations above. The architecture sits at what Newell called the “deliberate act level,” roughly 100 milliseconds, where elementary operations are selected and applied.
2.2 Human Associative Memory and Rational Analysis
ACT-R’s declarative memory stores facts as chunks—structured units with typed slots that encode relational information. These chunks are annotated with quantitative metadata, most importantly activation values that reflect the frequency and recency of their use. The activation equation combines a base-level component (reflecting the chunk’s history of use) with an associative component (reflecting the relevance of the current context). Retrieval from declarative memory is probabilistic: chunks with higher activation are more likely to be retrieved and are retrieved faster.
This is where Anderson’s methodology of rational analysis becomes central. Rather than treating the activation equation as an arbitrary curve-fitting exercise, Anderson derived it from a Bayesian analysis of the statistical structure of the environment. The base-level learning equation reflects the empirical finding that the probability of needing a piece of information follows a power-law decay from its last use—a regularity Anderson and Schooler (1991) documented across domains as diverse as newspaper headlines and child-directed speech. The architecture is thus adapted to its environment in a principled mathematical sense: its memory system implements an approximately optimal solution to the problem of retrieving information given the statistical structure of real-world demands.
This same rational-analysis methodology extends to ACT-R’s procedural system. Production rules are selected not just by symbolic matching but also by utility values that are learned through reinforcement. The utility learning mechanism adjusts the expected payoff of each production based on whether it contributed to achieving the system’s goals, implementing a form of temporal-difference learning. The result is that ACT-R gradually learns to select the most effective strategies through experience, without requiring explicit programming of preference orderings.
2.3 The Adaptive Control of Thought
The “adaptive” in ACT-R (Adaptive Control of Thought—Rational) is not ornamental. Anderson’s central theoretical claim is that human cognition is adapted to the statistical structure of the environment in a way that can be modeled as approximately Bayesian optimization under resource constraints. This is bounded rationality in Herbert Simon’s sense, but with the specific constraints determined by the cognitive architecture rather than left unspecified.
The procedural system operates through a recognize-act cycle. On each cycle, all production rules whose conditions match the current state of the buffers are identified (the matching phase). Among these, one is selected based on utility values and noise (the selection phase). The selected production then fires, modifying buffer contents and potentially initiating retrievals from declarative memory or actions in the external environment (the execution phase). Complex behavior arises from sequences of these cycles, each operating in its local context—there is no separate architectural module for global optimization or planning.
This last point is significant. Planning, language processing, analogical reasoning, and other complex cognitive activities are hypothesized to emerge from sequences of these primitive cognitive cycles, not from dedicated architectural modules. The architecture provides the fixed substrate; the flexibility comes from the knowledge and skills that are acquired and processed by that substrate. In Anderson’s terms, intelligent behavior arises from the combination of an implementation of a cognitive architecture plus knowledge and skills.
3. Carruthers’ Centered Mind: Consciousness, Working Memory, and the Sensory Base
3.1 The Sensory-Based Account of Thought
Peter Carruthers’ The Centered Mind develops a radical thesis about the nature of conscious thought that has deep implications for how we understand cognitive architecture. His central claim is that conscious thought is always sensory-based, relying on the resources of the working memory system. When we reflect—examining possibilities, considering options, making inferences, arriving at judgments—we do so by sustaining and manipulating sensory images (including inner speech, visual imagery, and other modality-specific representations) through attentional signals directed at midlevel sensory areas of the brain.
This challenges the philosophical default assumption that conscious thinking consists (at least partly) of amodal propositional attitudes—that when you consciously judge that it will rain tomorrow, the content of your consciousness includes the abstract, non-sensory proposition itself. Carruthers argues instead that amodal attitudes (beliefs, goals, decisions, intentions) are never conscious. They operate entirely in the background, “pulling the strings” by selecting, maintaining, and manipulating the sensory-based contents that do figure consciously in working memory. The conscious mind is, as Carruthers vividly puts it, somewhat like a marionette controlled by off-stage actors who do their work unseen.
3.2 Working Memory as the Theater of Consciousness
Carruthers builds his account on the cognitive science of working memory, particularly the framework associated with Baddeley and colleagues but significantly updated with neuroscientific findings. The classical model posited a central executive controlling two “slave” subsystems: a visuospatial sketchpad and a phonological loop. Contemporary evidence, however, reveals that working memory utilizes the resources of perceptual systems across all sense modalities—not just vision and audition, but also smell, touch, proprioception, and affect. Whenever brain imaging has been used to investigate working-memory tasks, activity in the relevant sensory area has been found.
This is a critical finding for cognitive architecture. It means that working memory is not a separate storage buffer but rather a mode of operation of the perceptual systems themselves, sustained by top-down attentional signals from prefrontal cortex. The same mechanisms used during online processing of visual input to categorize and conceptualize a stimulus are used offline to create neural activity in midlevel visual areas similar to what would occur if an appropriate object were being perceived. The same category-specific patterns of activity occur in both midlevel and higher-level areas of the temporal cortex visual-processing stream, regardless of whether instances of those categories are perceived or imagined.
This connects directly to the architecture question. If working memory is implemented by sustained activity in sensory cortices under executive control, then the “working memory” component of cognitive architectures like ACT-R and the Standard Model is not a single homogeneous buffer but a distributed system whose contents are inherently sensory in character. The “global workspace” through which modules communicate is implemented by the very perceptual machinery that processes incoming information from the world.
3.3 Attention as the Mechanism of Consciousness
Carruthers draws on global workspace theory (Baars, Dehaene) to argue that access-conscious states are those whose contents are globally broadcast to a wide range of consumer systems—memory encoding, emotional evaluation, decision-making, verbal report, and so on. The mechanism that determines which representations achieve this global broadcasting is attention. Attended sensory representations in working memory become conscious; unattended ones do not.
But what is attention? Carruthers argues that attending is a form of action—specifically, a form of mental action that involves directing processing resources toward particular representations. This is not metaphorical. The neural circuits involved in directing attention overlap substantially with those involved in planning and executing overt motor actions. The frontal eye fields, for instance, are involved in both directing eye movements and directing covert visual attention. This action-based conception of attention has a crucial implication: the stream of consciousness is thoroughly active in nature. Even mind-wandering, which subjectively seems passive, results from unconscious cost-benefit decisions to redirect attention.
3.4 Dual Systems and the Architecture of Reasoning
Carruthers engages extensively with dual-process theories of reasoning (System 1 / System 2 in Kahneman’s popularization). System 1 processes are fast, automatic, parallel, and unconscious. System 2 processes are slow, controlled, serial, and conscious. On Carruthers’ account, System 2 reasoning just is reasoning conducted through the working memory system—manipulating sensory-based representations under executive control, cycling through sequences of attended states.
This provides a natural bridge to the cognitive architecture literature. The “serial bottleneck” identified in architectures like ACT-R and Soar—the constraint that only one deliberate act can be selected per cognitive cycle—corresponds to the capacity limitations of System 2 reasoning. System 1 processes, meanwhile, correspond to the massively parallel processing that occurs within modules (perceptual processing, memory retrieval, motor preparation) between cognitive cycles. The architecture’s serial core operating over parallel periphery maps neatly onto the dual-systems framework, but with the architectural account providing the mechanistic detail that dual-systems theory typically leaves unspecified.
Carruthers raises a provocative question about whether System 2 reasoning constitutes a “second mind” or an “extended mind.” He argues against both. System 2 is not a separate mind but rather the same cognitive architecture operating in a particular mode—one that routes processing through the working memory bottleneck rather than allowing it to proceed entirely within individual modules. And while external representations (notes, diagrams, calculators) can scaffold this process, the core reasoning mechanism remains internal.
4. The Standard Model of the Mind: Toward Consensus
4.1 Why a Standard Model?
The 2017 paper by Laird, Lebiere, and Rosenbloom proposes developing a standard model of humanlike minds, analogous to the Standard Model of particle physics. The motivation is that decades of work on cognitive architectures—each developed largely independently, with different goals, terminologies, and implementation choices—have converged on a set of shared assumptions that had never been made explicit. Making this consensus visible could accelerate progress by providing a common reference point, facilitating comparison across architectures, and guiding both researchers and practitioners.
The proposed standard model is not itself an implemented architecture. It is a theoretical architecture: a statement of what must be in a cognitive architecture in order to provide a humanlike mind. The intent is not unanimity but consensus—a best current understanding that will inevitably be incomplete and will need revision as more is learned. Omission from the standard model is often a statement of where consensus is needed, rather than a consensus on lack of importance.
4.2 Core Structure and Processing
The Standard Model posits that the mind is built from independent modules with distinct functionalities. The core components are:
Perception — converts external signals into internal representations with associated metadata, placing results in specific working memory buffers. There can be multiple perceptual modules for different modalities, each with an attentional bottleneck constraining what reaches working memory. Perception can be influenced by top-down expectations from working memory.
Working Memory — a temporary global space within which representations can be dynamically composed from the outputs of perception and long-term memories. It includes buffers for initiating retrievals and motor actions, maintaining intermediate results, and tracking goals. All of working memory is available for inspection and modification by procedural memory.
Declarative Long-Term Memory — a persistent store for facts and concepts, structured as a graph of symbolic relations annotated with quantitative metadata (frequency, recency, co-occurrence, similarity). Retrieval is cue-based: a cue is created in a working memory buffer, and the result is deposited back in that buffer.
Procedural Long-Term Memory — contains knowledge about actions, both internal and external. Based on pattern-directed invocation: rule-like conditions match against working memory contents, and a single deliberate act is selected per cycle, with metadata influencing the selection. This is the locus of global control.
Motor — converts symbolic structures in its buffers into external action through control of effectors.
The heart of the model is the cognitive cycle. Procedural memory tests working memory, selects a single deliberate act, and executes it by modifying working memory. These modifications can initiate further processing: memory retrievals, motor actions, top-down influence on perception. Complex behavior—both external and internal—arises from sequences of such cycles, operating at roughly 50 ms per cycle in human cognition. There is no separate module for global planning or optimization; all complex cognitive activities (planning, language processing, Theory of Mind) emerge from sequences of primitive acts composed through the cognitive cycle.
4.3 Memory and Content: The Hybrid Representation
A crucial feature of the Standard Model is the hybrid of symbolic and subsymbolic processing. This represents perhaps the most dramatic evolution from the early days of purely symbolic cognitive architectures. The core content is represented as relations over symbols—the primitive elements over which relational structures (semantic networks, ontologies, taxonomies) can be defined. But supplementing these symbolic structures is quantitative metadata that annotates symbols and relations: frequency, recency, activation, utility, similarity, and other numeric values.
There is a strict distinction between domain data (symbols and relations) and metadata (numeric annotations). Metadata exists only in support of symbolic representations, and relations cannot be defined over metadata. The available metadata types and associated processing mechanisms are fixed within the architecture. This is what yields the “subsymbolic” or statistical character of modern cognitive architectures—not the replacement of symbols by vectors (as in neural networks) but the annotation of symbols with numeric information that modulates how they are stored, retrieved, and selected.
The Standard Model is agnostic about whether symbols are implemented as uninterpreted labels (as in Lisp, Soar, and ACT-R), as patterns over distributed vectors (as in Spaun’s semantic pointers), or as both (as in Clarion and Sigma). What matters is that they provide the necessary functionality to represent and manipulate relational structures—addressing the binding problem of how multiple elements can be associated in structured ways.
4.4 Learning
All types of long-term knowledge are assumed to be learnable. Learning is incremental, online, and occurs as a side effect of performance—typically based on some form of backward flow of information through internal representations. Procedural memory involves at least two learning mechanisms: one that creates new rules from the composition of rule firings (yielding behavioral automatization, as in Soar’s chunking or ACT-R’s production compilation) and one that tunes action selection through reinforcement learning. Declarative memory similarly involves mechanisms for creating new relations and tuning associated metadata.
More complex forms of learning—such as acquiring entirely new cognitive strategies or restructuring knowledge—arise from combinations of these fixed, simpler learning mechanisms. Learning over longer timescales accumulates from shorter-term learning episodes, and can include explicit deliberation over past experiences.
4.5 The Convergence
The paper includes a striking analysis showing how ACT-R and Soar—which started from rather different bases—have converged dramatically over the decades. By 2016, both architectures are in near-total theoretical agreement with the Standard Model, differing mainly in the extent to which they implement perception and motor systems. Sigma, being newer, shows more incompleteness but also more disagreement, particularly in its claim that capabilities the other architectures implement through dedicated modules can instead emerge from interactions among more primitive mechanisms. This disagreement is itself illuminating: it points to a genuine open question about how much modularity is architecturally fixed versus arising from knowledge and experience.
5. Rosenbloom’s Architectural Exploration: From Soar Through Sigma
5.1 The XAPS Prelude and the Power Law of Practice
Rosenbloom’s memoir reveals that his career-long pursuit of cognitive architectures began with an attempt to fuse symbolic rule-based processing with neural-like activation flow in the late 1970s. The XAPS (eXperimental Activation-based Production System) architectures explored what happens when rules guide the flow of activation associated with symbol structures. One crucial lesson emerged: unconstrained activation processing is a morass. Making it work requires either constraining the processing to match probability theory (as in graphical models) or taming it through learning algorithms that align it with experience (as in backpropagation). This insight would not bear full fruit for three decades, until Sigma.
The work on the power law of practice, conducted with Newell, established that human performance improvement follows a power-law function across an extraordinary range of tasks. The explanation they developed—chunking, combining multiple processing steps into a single compiled step—became a domain-independent architectural learning mechanism. Making chunking impasse-driven (triggered by gaps in knowledge needed for decision-making) yielded a simple, elegant solution that became central to Soar.
5.2 Soar: Elegant Unification
Soar combined rules, problem spaces, impasse-driven subgoaling, and chunking into a remarkably parsimonious architecture: one representation for long-term knowledge (rules), one way to make decisions (symbolic preferences), one way to learn (chunking), and one way to reflect on its own behavior (impasse-driven subgoaling). Yet from this small set of mechanisms, a wide range of critical behaviors emerged.
The key principle was listening to the architecture—Newell’s maxim that the architecture itself should guide how new capabilities are implemented. Rather than adding a dedicated module for each new capability (planning, declarative memory, various forms of learning), the Soar team explored whether existing mechanisms could produce the desired behavior through novel interactions. Data chunking, for instance, emerged as an unexpected form of declarative learning from the interaction between search and chunking—what had been designed as a speed-up mechanism turned out to support entirely new kinds of learning. Planning could arise without explicit planning modules, with chunking storing plans piecemeal into rule memory.
Soar was applied to military simulation (Intelligent Automated Forces), natural language processing, robotics, and modeling of human cognition. By the early 1990s, Newell used it as the basis for his William James Lectures arguing for Unified Theories of Cognition. But by the late 1990s, Rosenbloom felt he had reached a dead end: he could not see how to extend Soar further while maintaining the constraint of simplicity and elegance. John Laird continued developing Soar by adding new memories and learning mechanisms (semantic memory, episodic memory, reinforcement learning, mental imagery), yielding a more capable but also more modular architecture. Rosenbloom stepped away entirely for a decade.
5.3 Sigma: Grand Unification Through Graphical Models
The breakthrough that launched Sigma came from recognizing that probabilistic graphical models—specifically, factor graphs processed by the sum-product algorithm—could provide a single computational substrate general enough to subsume both symbolic rule-based processing and statistical/neural processing. Where Soar’s rule-based memory was replaced by an extension of statistical-relational factor graphs, and Soar’s chunking mechanism was replaced by gradient descent learning.
Sigma crystallized around four desiderata: grand unification (spanning high-level symbolic thought and low-level perception/motor processes without separate modules), generic cognition (modeling both natural and artificial humanlike minds), functional elegance (generating broad capabilities from interactions among a small set of general mechanisms), and sufficient efficiency (running fast enough for real applications).
The result was an architecture with a single long-term memory general enough to support rule-based procedural processing, semantic and episodic declarative memories, probabilistic reasoning, neural networks, and forms of perception and motor control—all without dedicated modules for each. Two architectural layers (one graphical, one cognitive) plus specialized knowledge and skills yielded the range of capabilities that other architectures achieve through proliferating modules.
Rosenbloom describes this as deconstruction: taking capabilities traditionally implemented as distinct modules (reinforcement learning, episodic memory, mental imagery) and reinterpreting them as combinations of more basic mechanisms. These reinterpretations provided some of the strongest intellectual breakthroughs of his career—moments when a deeper understanding seemed to emerge from taking a new perspective on familiar capabilities. The experience convinced him that functional elegance is not merely an aesthetic preference but a genuine source of scientific insight.
5.4 Emotion as Architectural Modulation
One of the most philosophically consequential developments in Sigma is its treatment of emotion — not as a dedicated affective module bolted onto the side of cognition, but as a pervasive modulation arising from the architecture’s own computational operations. This follows from the deconstruction principle: if all cognition is inference on a factor graph, then emotion should emerge from the same substrate rather than requiring special-purpose machinery.
Sigma derives emotional states from two primary appraisal variables that are computed continuously and everywhere in the graph. Expectedness measures the divergence between the architecture’s predictive models and incoming observations — when forward inference on the graph generates a prediction that the current state violates, expectedness drops, and the architecture registers surprise. Desirability tracks the gradient of the goal state — whether the current configuration of state variables reduces or increases distance to the agent’s objectives. These are not post-hoc annotations applied to completed inferences. They are intrinsic byproducts of the graph’s ongoing message-passing, computed at every factor node as part of the same sum-product algorithm that handles reasoning, retrieval, and motor control.
The architectural consequences are immediate. Low expectedness automatically reallocates processing resources toward the discrepant input — a functional analogue of attentional capture by surprising stimuli. The interaction of desirability and controllability biases action selection: negative desirability combined with low controllability favors conservative, well-rehearsed responses, while positive affect widens exploration. Crucially, the gradient-descent learning mechanism that operates over the graph also refines these appraisals over time — the agent effectively learns what to feel in response to recurring situations, adjusting its emotional responses based on accumulated reward signals.
This implementation has significant implications for the broader project of the Common Model of Cognition. A 2024 proposal for extending the Common Model to emotion, heavily influenced by Sigma’s approach, introduces a Metacognitive Assessment Module that continuously evaluates working memory contents against long-term goals and homeostatic drives, generating high-dimensional appraisal vectors that modulate global architectural parameters — attention bandwidth, retrieval strategy bias, learning rate, action-selection temperature. The key move, both in Sigma and in the proposed extension, is that emotion is not a feeling appended to cognition but a control signal that is cognition operating in its evaluative mode.
This resonates with a point we will need when we turn to LLM interpretability. The “motley mix” finding — that transformer models simultaneously deploy principled reasoning circuits alongside shallow heuristics, with semantic content sometimes overriding formal logic — describes a system that lacks precisely this kind of global evaluative modulation. A transformer’s residual stream has no architectural mechanism for computing surprise at its own predictions or tracking desirability relative to persistent goals. The appraisal-based view suggests that what such systems are missing is not emotion in the folk-psychological sense but the homeostatic self-regulation that emotion implements: the capacity to monitor the quality of one’s own processing and adjust resource allocation accordingly. Whether the emerging “ensemble architectures” that wrap LLMs in explicit goal-tracking and memory-persistence layers can approximate this function is an open question — but it is a question that the cognitive architecture tradition, via Sigma, has given precise computational form.
5.5 The Common Model of Cognition
The Standard Model paper of 2017 was later renamed the Common Model of Cognition after community feedback that “Standard Model” sounded overly prescriptive. The Common Model represents a consensus theoretical architecture—a meta-architecture about the space of humanlike architectures rather than a single implementation. It has been mapped onto more than a dozen cognitive architectures and tested as a model of high-level brain architecture by Andrea Stocco’s team, who used functional brain scans to compare it with traditional hub-and-spoke and hierarchical proposals from neuroscience. The Common Model showed overwhelming statistical dominance over these alternatives.
The ongoing work includes extensions to emotion, metacognition, language processing, and the neural underpinnings of the model’s components. The social process of achieving community consensus has proven as challenging as the scientific content—working groups have gone quiescent, and the core team is rethinking how to maintain momentum. But Rosenbloom believes the Common Model may ultimately have a larger and more enduring impact than any of the individual cognitive architectures from which it was distilled.
5.6 Dichotomic Maps: Understanding the Space of Architectures
Rosenbloom’s most recent theoretical contribution is the development of dichotomic maps—theoretical architectures that leverage cross products of dichotomies (this-or-that distinctions) to structure and understand spaces of cognitive/AI technologies. Three core dichotomies have been identified: (a)symmetric (whether processing proceeds in one direction or uniformly in any direction), (un)controlled (whether decisions are required during processing), and (meta)data (whether representations are symbolic data or quantitative metadata).
These maps serve two purposes. Disciplinary maps structure the technologies underlying AI and cognitive science, revealing that many traditional dichotomies (rules vs. logic, procedural vs. declarative, feedforward vs. autoassociative networks, supervised vs. unsupervised learning) are instances of the same core (a)symmetric distinction. Architectural maps show which technologies are combined within a specific architecture, revealing gaps and structural similarities. For instance, mapping AlphaZero onto this framework reveals it as surprisingly similar to Soar in its core structure, differing mainly in lacking a symmetric uncontrolled (reasoning/declarative) capability and in using neural rather than symbolic representations.
The deeper insight is that rather than forcing a choice between opposing ends of these dichotomies, the best architectures span them—incorporating both asymmetric and symmetric processing, both controlled and uncontrolled mechanisms, both symbolic and subsymbolic representations—to achieve the complementary benefits of each.
6. Cross-Cutting Themes
6.1 The Serial Bottleneck
All four works converge on the existence of a serial bottleneck at the heart of cognition. In ACT-R, only one production fires per cycle. In Soar, a single deliberate act is selected per cognitive cycle. The Standard Model explicitly posits this constraint. And Carruthers’ analysis of working memory and System 2 reasoning identifies the same bottleneck from a philosophical and psychological perspective: the capacity limitations of conscious, controlled thought reflect the architectural constraint that only one act of attention or executive control can be executed at a time.
Yet this serial core operates within a massively parallel context. Processing within modules proceeds in parallel (perceptual recognition, memory retrieval, motor preparation all happen simultaneously). Multiple modules operate asynchronously. The serial bottleneck in procedural memory’s interaction with working memory provides just enough seriality for coherent, goal-directed thought while preserving the efficiency of parallel processing everywhere else.
6.2 The Symbolic-Subsymbolic Hybrid
The purely symbolic cognitive architectures of the 1970s and 1980s have given way to hybrid systems that integrate symbolic structures with quantitative metadata. ACT-R pioneered this with activation-based retrieval and utility-based selection. Sigma takes it furthest by grounding the entire architecture in factor graphs that inherently blend symbolic and probabilistic computation. The Standard Model makes it official: the hybrid of symbolic data and quantitative metadata is a consensus feature, not an idiosyncratic choice.
Carruthers’ analysis adds a philosophical dimension to this: the contents of conscious working memory are sensory representations—inherently graded, analog, metric—bound to abstract conceptual representations. The symbolic content (the propositional attitude) is never itself conscious; what is conscious is always the sensory vehicle. This suggests that the symbolic-subsymbolic distinction in cognitive architectures may correspond to a deeper distinction between the abstract computational content of thought and the sensory medium through which it becomes available to consciousness.
6.3 Working Memory as the Hub
Working memory emerges from all four perspectives as the central hub of cognitive processing. In the Standard Model, it acts as the inter-component communication buffer, the only place where information from different modules can be brought together and acted upon by procedural memory. In ACT-R, the buffers associated with each module constitute the aggregate working memory through which the procedural module exerts control. In Carruthers’ account, working memory is the theater of consciousness itself—the global workspace through which representations achieve the broadcasting needed for conscious access. And in Soar and Sigma, working memory is the substrate that the cognitive cycle operates over.
But the accounts differ on the nature of working memory. Carruthers emphasizes that it is implemented by sustained activity in sensory cortices, making it inherently modality-specific. ACT-R models it as a set of module-specific buffers. The Standard Model allows it to be seen as unitary or as separate modality-specific memories. These differences may reflect genuine uncertainty about the level of architectural commitment warranted by current evidence.
6.4 Learning as a Side Effect of Performance
A striking consensus across the architectures is that learning occurs incrementally and online, as a side effect of performance—not as a separate phase requiring distinct processing. Soar’s chunking creates new rules during problem-solving. ACT-R’s utility learning adjusts production strengths based on goal achievement during task performance. Sigma’s gradient descent operates continuously. The Standard Model codifies this as a general principle.
This contrasts sharply with the dominant paradigm in machine learning, where training and inference are typically distinct phases with different algorithms and even different hardware. The cognitive architecture tradition suggests that truly humanlike learning must be integrated into the performance loop—learning from every cognitive cycle, not just from curated training sets.
6.5 Bounded Rationality as the Organizing Principle
The Standard Model explicitly identifies bounded rationality—Herbert Simon’s concept—as the purpose of architectural processing, in contrast to the optimality that dominates much of AI. The architecture is not designed to find globally optimal solutions but to support effective decision-making under the constraints of limited memory, limited time, and limited knowledge. Anderson’s rational analysis shows that ACT-R’s specific constraints can be derived as approximately Bayesian-optimal solutions to information retrieval under realistic environmental statistics. The architecture is adapted to its niche, not abstractly optimal.
Carruthers’ analysis of how System 2 reasoning works through working memory manipulation adds a mechanistic account of why rationality is bounded: the serial bottleneck and capacity limitations of working memory impose fundamental constraints on the depth and breadth of deliberation. We reason by cycling through sequences of attended sensory representations, each constrained by what can be held in working memory at once. The “bounds” on bounded rationality are architectural features, not accidental limitations.
7. Open Questions and Future Directions
7.1 Consciousness and Architecture
Perhaps the deepest open question is the relationship between cognitive architecture and consciousness. The Standard Model is silent on phenomenal consciousness. ACT-R treats consciousness functionally, in terms of what information is available in the buffers. Carruthers offers the most developed account: access consciousness is global broadcasting of attended sensory representations in working memory, and phenomenal consciousness reduces to (or is constituted by) this access consciousness. But whether any cognitive architecture can truly explain consciousness rather than merely correlating with it remains contested.
The sensory-based account suggests a provocative architectural constraint: if conscious thought requires sensory vehicles, then architectures that process only abstract symbol structures may be inherently incapable of supporting consciousness in the relevant sense. This would mean that the question of machine consciousness is partly an architectural question—about whether the system’s representations have the right kind of sensory grounding.
7.2 The Challenge from Large Language Models
The recent emergence of large language models (LLMs) poses a fascinating challenge to the cognitive architecture tradition. LLMs achieve remarkable generality through a single mechanism (transformer-based neural networks trained on vast text corpora) rather than through the modular, hybrid architectures described here. Rosenbloom notes that LLMs have begun to blur the distinction between narrow and general AI. The Standard Model paper’s aside about “no additional specialized architectural modules” for complex activities like language processing and Theory of Mind takes on new significance when LLMs appear to do both without any of the traditional architectural machinery.
Yet LLMs also have well-documented limitations that the cognitive architecture tradition predicted: difficulties with systematic compositionality, lack of persistent memory and learning from experience, inability to ground language in perception and action, and absence of the kind of goal-directed, deliberate reasoning that the serial bottleneck enables. The emerging consensus may be that LLMs and cognitive architectures need each other—with LLMs providing the statistical knowledge and flexible language processing that architectures have struggled with, and architectures providing the structured reasoning, persistent memory, and grounded action that LLMs lack.
7.3 The Space of All Possible Minds
Rosenbloom’s dichotomic maps point toward an even more ambitious question: understanding not just humanlike minds but the full space of possible minds—including animal minds, artificial minds of various kinds, and perhaps alien minds. The Common Model deliberately restricts itself to humanlike architectures, acknowledging that there may be fundamentally different ways to achieve human-level intelligence. But the dichotomic maps provide tools for reasoning about this broader space: by identifying the core dichotomies that structure cognitive technologies and the criteria that explain why each end is useful, they offer a principled way to explore which combinations of capabilities are possible, which are effective, and which are humanlike.
8. Hippocampal Systems and Complementary Learning
Carruthers treats the hippocampus as central to three capacities: binding episodic memories (using spatial and temporal coordinates to link sensory representations across cortical areas), supporting prospection (constructing and evaluating future scenarios by recombining episodic elements), and contributing to the default mode network’s stimulus-independent activity (the mind-wandering that fills much of waking life). A widely accepted generalization in cognitive neuroscience — that memories are stored where they are processed, within the same systems that gave rise to the information — means the sensory content of episodic memory lives in sensory cortex, while the hippocampus provides the binding that links these distributed traces into coherent episodes. When episodic memories are activated, they are globally broadcast in the same manner as perception, entering consciousness through the same working-memory machinery.
What Carruthers does not engage with is why the hippocampal system exists as a separate architectural component in the first place. McClelland, McNaughton, and O’Reilly’s Complementary Learning Systems (CLS) theory provides the computational answer. Neural networks with distributed representations face a fundamental tradeoff: rapid learning of new information causes catastrophic interference — new learning overwrites the statistical regularities painstakingly extracted from prior experience. The solution is two complementary systems with different learning rates. The neocortex learns slowly, through interleaved exposure, to extract the statistical structure that constitutes general knowledge (semantic memory, concepts, skills). The hippocampus learns rapidly from single episodes, storing sparse pattern-separated traces that can later be “replayed” to the neocortex — during sleep, rest, or idle moments — for gradual consolidation.
A 2016 update by Kumaran, Hassabis, and McClelland broadened the theory in important ways. Hippocampal replay is not merely passive re-presentation; it allows goal-dependent weighting of experience statistics, so the system can prioritize consolidating what matters. Furthermore, neocortical learning can be rapid when new information is consistent with known structure — schema-consistent learning benefits from existing scaffolding. The theory thus predicts an asymmetry: genuinely novel information must go through slow hippocampal-to-neocortical consolidation, while information that fits established schemas can be integrated quickly.
This directly addresses an unresolved question in the Standard Model. The 2017 paper notes that there is not yet a consensus on whether there is a single uniform declarative memory or two separate memories, one semantic and the other episodic. CLS theory provides the computational argument for why two are needed: they embody complementary learning rates that solve the stability-plasticity dilemma. The Standard Model’s semantic memory corresponds to the slow neocortical system that extracts cross-episode statistical structure; its episodic memory corresponds to the fast hippocampal system that stores individual experiences. They are not merely different contents within a single memory but different learning regimes with different computational properties.
The connection to concept formation is immediate. The slow neocortical learning system is essentially the substrate for the kind of concept acquisition described by prototype theory and the Bayesian program — gradually extracting the statistical regularities that define categories from interleaved exposure to instances. The fast hippocampal system provides the episodic grounding that exemplar theory and Prinz’s proxytype account emphasize — retaining specific encounters that can be consulted for context-sensitive categorization. Machery’s heterogeneity thesis — that “concepts” actually involve multiple distinct psychological structures deployed for different tasks — maps onto the CLS architecture: prototype-like processing draws on neocortical statistical summaries, while exemplar-like processing draws on hippocampal episodic traces. These are not competing theories of the same thing but descriptions of two complementary systems, each optimized for different computational demands.
For the cognitive architecture tradition, CLS theory also illuminates the learning principle that the Standard Model codifies — that all learning is incremental and online, occurring as a side effect of performance. This is true of both the fast and slow systems, but in different senses. Hippocampal encoding is rapid and online — it captures episodes as they happen. Neocortical consolidation is slow and online — it integrates replayed episodes during ongoing processing. Neither system has a separate “training phase” distinct from performance, which is precisely the architectural principle the Standard Model emphasizes. But the two systems implement this principle at different timescales and with different consequences for what kind of knowledge results.
9. Four Pointers and the Merge Conjecture
The empirical finding that working memory sustains approximately four chunks of information (Cowan 2001) is one of the most robust results in cognitive psychology. Activity in the intraparietal sulcus increases with memory load and plateaus at about four items. Visual and auditory working memory share the same attentional resources, with the maximum number of items retained in each domain matching what can be retained in cross-modal tasks. The limit appears to reflect attentional capacity rather than storage capacity — it is the number of representations that top-down attention can simultaneously sustain in an active, globally broadcast state.
Recent models suggest that this capacity may be implemented by a small number of pointer-like attentional indices — approximately two per hemisphere — each of which can latch onto a sensory representation in midlevel cortex and sustain it through top-down signals from prefrontal and parietal regions. These are functionally similar to the “object files” described in the vision literature (Kahneman and Treisman; Pylyshyn): location-based indexicals that anchor to particular entities, with folders attached into which property information is placed. In Carruthers’ framework, these pointers are the mechanism of the global workspace — what is pointed to is globally broadcast and thereby conscious; what loses its pointer decays.
If working memory consists of a small number of pointers to sensory representations, then the fundamental question becomes: what is the minimal operation these pointers can perform? We propose that the answer is Merge — the binary operation that takes what two pointers reference and combines them into a new structured object, which can then be re-attached to one of the pointers, freeing the other.
This is recognizably Chomsky’s Merge operation from the Minimalist Program: taking two syntactic objects A and B and forming the new object {A, B}. But we are proposing something stronger than the linguistic claim. Merge, on this view, is not specifically linguistic — it is the domain-general operation that the working memory architecture affords. It is what you inevitably get when you have a small number of pointers and the capacity to form structured combinations. The linguistic application of Merge — building hierarchical syntactic structures — is an exaptation of this more fundamental architectural operation.
Several converging lines of evidence support this conjecture. Fujita’s hypothesis of the motor-control origin of Merge argues that hierarchical syntax evolved from the capacity for sequential and hierarchical object manipulation — combining two actions into a structured sequence. This is directly compatible: the working memory pointers would provide the substrate for holding two action representations simultaneously while the motor system combines them. A February 2025 eLife paper from the Frank lab demonstrates a neural network model of prefrontal cortex and basal ganglia in which a ring attractor mechanism merges sensory representations with their nearest neighbor in PFC, and — critically — the resulting chunk can replace the original representation in a working memory slot. This is our proposal realized computationally: two items merge, the result re-attaches to one pointer, the other pointer is freed.
This reinterpretation also unifies the chunking traditions across cognitive architectures. In Soar, chunking compiles sequences of rule firings into new rules that accomplish in one step what previously required multiple steps. In ACT-R, chunk formation in declarative memory creates new structured units from experienced combinations. The Standard Model codifies procedural composition (yielding behavioral automatization) and declarative chunk creation as fundamental learning mechanisms. In every case, the underlying operation has the same structure: two or more elements are combined into a single structured unit that occupies one “slot” where multiple slots were previously required. This is Merge, operating at the level of working memory pointers.
The capacity limit of approximately four pointers then acquires an evolutionary interpretation. Four is the minimum number that supports recursive application of Merge with sufficient depth for useful structure-building. With two pointers, you can perform one Merge operation. With four, you can Merge two pairs independently and then Merge the results — yielding hierarchical structure with two levels of embedding. This corresponds closely to the practical depth of center-embedding in natural language (which rarely exceeds two levels despite unbounded recursive grammar) and to the depth of hierarchical planning in motor action. The brain did not evolve excess working memory capacity because it did not need to: four pointers suffice for the recursive Merge operations that natural language, action planning, and analogical reasoning require.
This conjecture connects the cognitive architecture tradition (working memory as the hub of the Standard Model) to Minimalist linguistics (Merge as the fundamental operation) to the learning literature (chunking as the mechanism by which experience is compressed into reusable units). It suggests that the serial bottleneck at the heart of cognitive architecture is not merely a constraint imposed by limited neural resources but a design feature — the minimal substrate for a binary combinatorial operation that, through recursive application and consolidation into long-term memory, generates the open-ended structured thought that distinguishes human cognition.
10. Conclusion
Architectures that began from different premises — ACT-R from cognitive psychology and rational analysis, Soar from AI and problem-solving, Sigma from probabilistic graphical models — have arrived at shared commitments: modular organization around a central working memory, a serial bottleneck driven by procedural memory, hybrid symbolic-subsymbolic representations, incremental online learning, and bounded rationality as the organizing principle. That this convergence can be captured in a Common Model that also outperforms neuroscientific alternatives as a model of high-level brain organization is significant. Carruthers’ philosophical analysis adds that even within this consensus, open questions remain about the nature of the representations involved and about the relationship between architectural processing and conscious experience.
The two extensions proposed in sections 8 and 9 are offered in the spirit of Rosenbloom’s architectural exploration — seeking insight through synthesis across disciplines. Grounding the semantic-episodic distinction in CLS theory’s complementary learning rates resolves an open question in the Standard Model and connects the architecture of memory to the architecture of concept formation. The conjecture that Merge over working memory pointers is the fundamental operation underlying both chunking and linguistic syntax connects the cognitive architecture tradition to Minimalist linguistics and to the learning literature, while offering a specific evolutionary rationale for the ~4-item capacity limit. Whether these connections hold up under closer scrutiny is a question for future work — but the convergence across fields suggests they are worth pursuing.
Rosenbloom’s memoir cites a maxim from Allen Newell: “Deep scientific ideas are exceedingly simple. Others usually see them as trivial.” The idea of a cognitive architecture — a fixed computational framework that enables and constrains the mind’s flexibility — is simple. Working out its details is not. It is, as Newell said, a question that can hold you for a lifetime.
This article was co-authored by Lukasz Stafiniak and Claude (Anthropic). The arguments, errors, and speculative leaps are jointly owned.
I think that chatbots have a wider range of emotions than any single human.