Cognitive Architectures: A Deep Dive into the Science of the Mind

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.

3.1 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.

4.1 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.

5.1 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 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.

5.5 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.

6.1 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.

7.1 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 Emotion, Motivation, and Personality

The Standard Model acknowledges significant incompleteness regarding emotion, motivation, and what Rosenbloom calls “personal” aspects of cognition. Rosenbloom’s Sigma project had begun to explore appraisal-based emotion and temporal motivation theory before his work was interrupted by illness. A 2024 proposal for extending the Common Model to emotion has been developed but not yet achieved consensus. This remains a major gap: any architecture claiming to model complete humanlike minds must account for the pervasive influence of affect on cognition, decision-making, and learning.

7.3 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.

7.4 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.