Cognitive Architectures

A cognitive architecture is a claim that the mind has fixed parts. Not fixed knowledge — fixed machinery: memories, a control cycle and a learning mechanism that stay the same whatever the task, the way an instruction set stays the same whatever the program. The field's own definition carries a double commitment, and it is the whole methodological argument: an architecture is “both a theory about the structure of the human mind and a computational instantiation of such a theory.” It has to be falsifiable as psychology and runnable as engineering, and the crisis of the last twenty years is that those two criteria came apart.

Allen Newell set the standard in Unified Theories of Cognition in 1990: not a microtheory per phenomenon but one structure meeting a list of constraints that runs from real-time operation and language through learning and development to self-awareness and neural realisability. Two long-running systems answered him — Soar, begun in 1983, and ACT-R, whose lineage starts with an associative-memory model in 1973 — and one spiking-neuron model, SPAUN, tried to satisfy the neural constraint directly with 2.5 million simulated neurons. The tradition then attempted a consensus, the Common Model of Cognition, which this brief must name without describing, for reasons given below.

Two claims are argued here. The first is that the constraint which has actually bounded this tradition for forty years is the utility problem — the cost of deciding which rules apply growing faster than the value of having more rules — and that this is an indexing problem, sitting in a field (databases, search, retrieval) that has been transformed since it was diagnosed, with no published check on whether it still binds. The second is that the planner-plus-memory-plus-tools-plus-reflection scaffolds now built around language models are recognisably the architecture the symbolic tradition specified, rebuilt largely without reference to it. Whether that is convergent design or a field forgetting its own literature is a real question, and this brief asks it rather than answering it.

Established Newell asked for something specific, and the request has never been withdrawn. Unified Theories of Cognition (1990) rejected the accumulation of single-phenomenon models as a route to a science of mind; Newell “didn't consider the construction of single simulative microtheories a sufficient mean” for unification. Established The constraints a unified theory had to satisfy, as enumerated in the reference account read for this brief, are: flexible adaptation to the environment; goal-oriented rational behaviour; real-time operation in complex environments; symbol and abstraction manipulation; natural and artificial language; learning and development; autonomous operation in social contexts; self-awareness; and neural realisability with evolutionary emergence. Established Item eight is self-awareness and item nine is the connectionist constraint — so the list that founded the symbolic programme already contained both the hard problem of consciousness and the demand that the theory be implementable in neurons. Frontier No architecture has ever addressed all nine, and the two hardest are the two the field quietly dropped. Established Underneath the list sits the physical symbol system hypothesis, stated by Newell and Simon in exactly these words: “A physical symbol system has the necessary and sufficient means for general intelligent action.” Established They received the Turing Award in 1975, not the 1976 that is usually printed — 1976 is when the lecture was published.

Established Soar is the programme's purest instance and it is still shipping code. It began in 1983 as John Laird's Carnegie Mellon thesis with Newell and Paul Rosenbloom, aiming at “fixed computational building blocks necessary for general intelligent agents.” Established Its components are the tradition's vocabulary in one place: a production system pairing procedural with working memory; the Problem Space Hypothesis, which treats behaviour as search through state spaces; a State-Operator-Result decision cycle; impasses, which automatically create substates when knowledge does not determine what to do next; chunking, which compiles the resolution of a substate back into a rule; a semantic memory with activation-based retrieval; an episodic memory that “automatically records snapshots of working memory in a temporal stream”; and reinforcement learning tuning rule preferences. Established It is C and C++ under a BSD licence, with a Java reimplementation, JSoar, maintained by SoarTech; the group is now the Center for Integrated Cognition, under Laird and Robert Wray. Established The project's own site announces Soar 9.6.5, introducing a test-literalization operator for finer control over chunking, migrating the build to CMake, and shipping SWIG bindings for Python, Java, Tcl, C# and JavaScript alongside chunking, episodic-memory and semantic-memory fixes. Frontier That release note is the strongest single piece of evidence against the abandoned-programme reading: it describes substantive work on learning semantics, not maintenance. Established No release date appears on the page, which is a small illustration of how thinly this field documents itself.

Established Soar's scale record is real, unambiguous, and thirty years old. TacAir-Soar and RWA-Soar flew fixed-wing and rotary-wing tactical air missions in DARPA's Synthetic Theater of War exercise, STOW-97, described as the largest fielding of synthetic agents in a joint battlespace over a 48-hour period. Established Others on the record: NL-Soar for natural language, NTD-Soar modelling NASA Shuttle coordination, game agents, robotic control and dialogue-capable virtual humans. Frontier Nothing since STOW-97 is larger, and a tradition whose flagship deployment is an exercise from 1997 is making a claim about its own trajectory whether it intends to or not. Established The applied literature has not stopped — a 2022 IEEE paper integrates Soar with a coherent radar model for target detection — but it exists here as a verified bibliographic record, not a read text. Frontier Thin, not dead.

Established ACT-R makes the stronger scientific claim and is in a visibly different state of activity. Its lineage runs from HAM, the Human Associative Memory model of Anderson and Bower in 1973, through ACT and ACT*, to ACT-R in 1993, when John Anderson incorporated rational analysis; 4.0 added perceptual-motor capability, 5.0 modules mapped to brain systems and the buffers connecting them, and 7.0 (2015) a remote interface with Python support. Established The current version is 7.21.6, released 21 December 2020. Established Architecturally it has two module classes — perceptual-motor, of which the visual and manual are most developed, and memory, split into declarative facts and procedural rules — joined by buffers, “specialized structures holding temporarily active information, reflecting cortical activity,” which are the only means of accessing modules apart from procedural memory. Established Anderson credits Newell as his major influence. Frontier The widely quoted production-cycle constant and the subsymbolic activation and utility equations are not printed here: the primary ACT-R site was behind an access gate during the research for this brief, and a number that everyone knows is still a number that has to be read somewhere.

Established ACT-R's real contribution is a validation method, and it is under-appreciated outside the field. Models produce a “step-by-step simulation of human behavior” yielding “quantitative predictions of latencies and accuracies” tested against behavioural data trial by trial — a timing prediction per step, not a fit to an aggregate curve. Established The architecture has further been extended to predict brain activation patterns, including BOLD response shapes in motor cortex, prefrontal cortex, anterior cingulate cortex and basal ganglia, by mapping buffers to regions. Frontier That is the strongest claim any cognitive architecture makes to being a theory of the brain rather than a theory of behaviour, and it is almost never mentioned when the tradition is dismissed. Established ACT-R models appear in more than 700 scientific publications, across the fan effect, primacy and recency, Stroop, task switching, language, Tower of Hanoi, algebra, driving, flying and menu selection. Frontier That list is also the case for the prosecution: laboratory microtasks, which is precisely what Newell's own complaint — that you cannot play twenty questions with nature and win — was aimed at, and unification was supposed to be the escape.

Established The field is small and it is not extinct, and its own census says so. The survey at arXiv:1610.08602 covers 84 cognitive architectures, of which 49 are actively developed, over 40 years, and discusses “over 900 practical projects implemented using” them, organised around core cognitive abilities: perception, attention, action selection, memory, learning and reasoning. Established The survey is attributed to Kotseruba and Tsotsos; its abstract page returned those figures but not its own title, author line or venue, so this brief cites it by identifier and asserts no journal name. Established The standard taxonomy divides the field into symbolic, connectionist and hybrid architectures, the named hybrid exemplar being CLARION — Ron Sun's system, an explicit top level of localist chunk nodes over an implicit bottom level of distributed microfeature nodes, with four subsystems: action-centred, non-action-centred, motivational and meta-cognitive, validated on serial reaction time, artificial grammar learning and process control. Established The lineage runs back past Soar to EPAM, out of Ed Feigenbaum's 1960 thesis. Speculative CLARION matters most here because a metacognitive subsystem and an implicit-explicit distinction are exactly the two things a language-model agent conspicuously lacks.

Established SPAUN is the neural-implementation route's one landmark, and it is smaller than its reputation. Chris Eliasmith's group built a model of 2.5 million simulated spiking neurons in structures resembling prefrontal cortex, basal ganglia and thalamus, implemented in Nengo on the Semantic Pointer Architecture, published as Eliasmith, Stewart, Choo, Bekolay, DeWolf, Tang and Rasmussen, “A large-scale model of the functioning brain,” Science 338(6111):1202-1205, 2012. Established Two independent reference sources give the neuron count and the citation identically, and both enumerate four capability areas: recognising numbers, remembering them, working out numeric sequences, and drawing numbers with a simulated arm. Established Neither gives a task count, a simulation speed, the hardware, a successor version or any criticism, and the Science paper was behind a paywall throughout the research for this brief. Frontier The commonly circulated figure of eight tasks is therefore not stated here as a fact. Established For scale in the other direction: the only quantified neuromorphic system in the sources read here is Stanford's Neurogrid, 16 NeuroCore chips emulating 65,536 neurons in 2014; TrueNorth and Loihi (2017) appear without neuron counts and the recent large systems do not appear at all. Frontier Those two figures are the ones this brief can defend, and the gap between them is the neural route's engineering problem in miniature.

Established The symbolic/subsymbolic split has a dating that undercuts the story usually told about it. Symbolic AI dominated from the mid-1950s to the mid-1990s, opening with the Logic Theorist in 1955-56; the first AI winter followed unmet promises on machine translation and autonomous vehicles by the mid-1960s; the expert-systems boom ran 1969 to 1986, with DEC's XCON as its flagship, cutting computer-configuration time by a ratio the source states implausibly enough that only the ratio is repeated here; and the second AI winter ran 1988 to 2011, driven by maintenance cost, clinicians' unwillingness to trust machine diagnoses, and specialist Lisp hardware losing to Unix workstations. Established Connectionism ran a mirror-image cycle: McCulloch and Pitts in 1943, Rosenblatt's perceptron in 1958, the first wave closed by Perceptrons in 1969; Hopfield in 1982 and backpropagation in 1986 opening the second; and a consensus that neural networks were “hopeless” holding until 2012, when GPUs arrived. Established The date the reference source gives for the Parallel Distributed Processing volumes conflicts with the conventional one, and this brief prints neither. Frontier Both traditions spent long stretches in disgrace, which is why the standard story is a sentence about the last fifteen years read backwards over seventy.

Established The sharpest objection to the symbolic programme was never about implementation, and the sharpest reply conceded more than is usually noticed. Fodor and Pylyshyn's 1988 challenge was systematicity: whether connectionist networks can explain the structured, compositional character of thought without classical symbolic representations. Established Smolensky's Integrated Connectionist/Symbolic architecture answered with tensor product variable binding, letting distributed systems manipulate symbolic structures, and the position that connectionist models can implement symbol manipulation comparable to the classical kind is now on the record. Established The reference account of connectionism says the framing has shifted: Neural Turing Machines (2014) and Relational Networks demonstrate that “connectionism and computationalism need not be at odds.” Established The named critics said different things and are routinely merged: Dreyfus (1979), that expertise rests on unconscious instincts not capturable in formal rules; Brooks (1990), that symbols are unnecessary because “the world is its own best model — it is always exactly up to date”; Searle, that the hypothesis addresses behaviour only and is irrelevant to consciousness; Nilsson, that treating symbol grounding as a requirement is erroneous. Established Hinton in 2016 called further investment in symbol manipulation “a huge mistake”; Marcus calls the hostility a “historical grudge” rather than evidence. Speculative What survives all this is a dispute not about what can be expressed but about what can be learned: whether structured representations can be acquired by gradient descent from realistic data. That is empirical, and large language models are the experiment.

Established The tradition has exactly one commercial artefact at scale, and its evidence base is honest and modest. Cognitive Tutor, guided by ACT-R, uses model tracing — checking every action a student performs against the architecture's model of the skill — and knowledge tracing, a Bayesian mastery estimate shown to the student on a Skillometer. It reached 75 US schools by 1999, and Carnegie Learning now supplies tutors to thousands. Established A 2013 Institute of Education Sciences review examined 27 studies, of which 6 met rigorous standards: 5 positive, 1 statistically significantly negative; a 2009 review of Cognitive Tutor Algebra I found potentially positive effects in 1 qualifying study of 14. Frontier That is the only place in this topic where an architecture has been evaluated by an outside body under a published methodology, and it is worth more than a larger claim.

Established The architectures came back as a framework for language-model agents, explicitly and by name. Sumers, Yao, Narasimhan and Griffiths published “Cognitive Architectures for Language Agents” (CoALA, arXiv:2309.02427, September 2023) in Transactions on Machine Learning Research. Established Its framework has three parts — “modular memory components, a structured action space to interact with internal memory and external environments, and a generalized decision-making process to choose actions” — and it explicitly connects contemporary language-agent research to classical cognitive science and symbolic AI. Frontier Set that against Soar's procedural, semantic and episodic memories, its structured operator space and its decision cycle, or ACT-R's modules and buffers, and the correspondence is not an analogy. It is the same decomposition. Speculative What that means is the open question. Convergent design would be evidence that the decomposition is forced — that anything acting over time under bounded resources needs these parts — which is close to what Newell claimed. Rediscovery without citation would be evidence about a field that discards its literature every fifteen years. Speculative The two are distinguishable, and the test is available: whether the parts these frameworks converged on are the parts the tradition specified, or a subset chosen by what is easy to build around a language model. Nobody has run it.

Frontier The most direct challenge to the whole programme is a fitted model that beats hand-built theories on the field's own criterion. Binz, Akata, Bethge and 38 co-authors released Centaur (arXiv:2410.20268, October 2024, revised April 2025), a foundation model trained on Psych-101: trial-by-trial data from “over 60,000 participants” performing “over 10,000,000 choices” in 160 experiments. Frontier The reported claims are better prediction of held-out participants than existing cognitive models, generalisation to novel scenarios, and internal representations more aligned with human neural activity after fine-tuning. Frontier Prediction of human behaviour is the criterion the architectures chose for themselves, and a fit is winning on it. Speculative The architectural reply is that Centaur predicts and does not explain, and that the dual definition makes explanation constitutive: an architecture must be a theory, and a large parameter fit is not one. Frontier Whether that objection is principled or defensive is this topic's live question, and it is answerable with a measurement nobody has taken — parameters spent per phenomenon explained, for both kinds of model, on the same data.

Established The two flagship architectures are in visibly different states and should stop being treated as one object. Soar ships releases with new learning semantics; ACT-R's most recent numbered version is dated December 2020. Frontier Neither fact is decisive alone — a stable architecture is what a mature theory looks like — but an account that says “the symbolic tradition” and means both will get the trajectory wrong in one direction or the other. Speculative What would settle it is funding, and that is precisely what this brief cannot report: the Soar project's about page and the primary ACT-R site were both behind access gates, so no budget, roadmap, headcount or grant appears anywhere here.

Frontier The field's own consensus attempt is named here and not described, and that gap is a finding about the field's visibility. The Common Model of Cognition — formerly the Standard Model of the Mind — is the Laird, Lebiere and Rosenbloom synthesis of what the major architectures agree on, and arguably the most important recent development inside the tradition. Established The research behind this brief obtained one verified bibliographic record for it — a chapter, “Common Model of Cognition: Searching for Consensus” in On Minds (MIT Press, 2026) — and no account of its content. Established No claim about the Common Model's content, structure or number of components appears anywhere in this brief. Speculative Notice what that implies: a synthesis by three of the field's most senior figures was, for a competent search under real access constraints, unreadable — a fact about how narrow this literature's public surface has become.

Frontier Neuro-symbolic integration has a taxonomy and no scoreboard. Kautz's six categories organise the space — Symbolic Neural (BERT, GPT-3); Symbolic[Neural] (AlphaGo's tree search with neural evaluation); Neural | Symbolic; Neural: Symbolic -> Neural, where symbolic reasoning generates training data; NeuralSymbolic, including Logic Tensor Networks (2016); and Neural[Symbolic], embedding symbolic reasoning inside networks — with named systems including Scallop, DeepProbLog, SymbolicAI, XNNs (2021), Abductive Learning (2022) and AllegroGraph (2023). Established Garcez and Lamb date the programme to the 1990s and frame it against Kahneman's System 1 and System 2; Marcus states the position as needing “the triumvirate of hybrid architecture, rich prior knowledge, and sophisticated reasoning.” Established The reference account of the field contains no quantitative benchmark results, and this brief prints none. Frontier A programme with six categories, a dozen named systems and no shared benchmark is in a specific and recognisable condition, and it is the condition the architectures were in for two decades.

Speculative The most exotic live possibility is that interpretability rebuilds the symbolic layer from the inside, vindicating the architectures rather than the architects. Mechanistic interpretability is a search for discrete, composable, nameable structures inside distributed systems — which is Smolensky's programme run in reverse. Speculative If the circuits recovered from a trained network prove compositional, reusable and capable of variable binding, then the physical symbol system hypothesis was right at the algorithmic level and wrong at the implementation level, and the tradition's error was insisting that symbols be written by hand. Handwave What would have to be true is precise: discovered circuits must be reusable rather than task-specific; they must support variable binding, the capacity Fodor and Pylyshyn demanded; and they must be causally manipulable the way a production rule is. Speculative The second is the binding condition and the sharpest unanswered question at the intersection of these two literatures — and essentially nobody poses it in these terms, because the interpretability community does not read the systematicity debate and the architecture community does not read circuits papers.

Established The target is an architecture that scales, and the three conjuncts are what make it hard. The impossible thing is a system whose performance improves monotonically with added knowledge and experience, without hand-authoring, while remaining inspectable as a theory. Established Drop inspectability and you have a transformer. Drop automatic acquisition and you have Soar in 1997. Drop monotonic improvement and you have every architecture ever built. Frontier The chain is short and mostly cheap: a like-for-like benchmark on which architectures and fitted models both compete; automatic knowledge acquisition into architectural form; retrieval that does not degrade as memory grows; inspectability preserved under acquisition; a working symbolic/subsymbolic interface; timing fidelity retained with a variable-latency component in the loop; impasse detection at scale; and monotonic improvement over months without an author editing the knowledge base. Established None of the eight is achieved; two have mechanisms waiting and no experiment.

Speculative The binding link is the utility problem, and naming it is the most useful thing this brief does. A production system decides what to do by matching its rules against working memory. As the rule set grows, the match cost per decision cycle grows with it, and past some point the system spends more time deciding than the extra rules save. Frontier This is the tradition's classic failure mode: Soar's chunking learns rules from resolved impasses, and historically the learning could make systems slower. The response was to bound what could be learned rather than to fix retrieval. Established No source read for this brief reports the problem solved, and none reports a measurement of it in a current release. Frontier That is the finding: the constraint is documented as having been binding, and then abandoned as a research target rather than resolved.

Speculative And the constraint is in the wrong field. Match cost as a function of rule-set size is an indexing problem: it asks how fast you can find the subset of stored patterns consistent with a changing state. That is the central question of database query planning, information retrieval, and approximate nearest-neighbour search — disciplines that have advanced enormously since the utility problem was diagnosed in the 1980s, and that routinely serve sublinear-time lookups over stores many orders of magnitude larger than any architecture's rule base. Speculative Nobody appears to have gone back to check whether the constraint that ended a research programme still binds against a modern index. Handwave If it does not, then a forty-year retreat was justified by a fact about 1980s data structures rather than a fact about cognition — which would be one of the cheapest and most embarrassing findings available in this category. Frontier The modern restatement is identical and equally unmeasured: a language-model agent with a growing episodic store faces the same curve, and current systems manage it by keeping the store small or the retrieval shallow, which is bounding learning under a new name.

Established This brief waits on two results produced elsewhere on the map. The first is measurement. Every comparison in this topic — architecture against fitted model, agent scaffold against end-to-end model, architecture against its own earlier version — is a benchmark claim, and the protocol questions that decide whether such a claim means anything belong to intelligence measurement: which split, how many attempts, at what compute, and whether the evaluation set existed publicly before the system was built. Established The specific instrument this topic needs from there is free-parameter accounting: a comparison of a hand-built theory against a fit is meaningless unless the parameters each spends are counted on the same basis. Frontier The second is the multi-agent case. If control structure is what architectures supply, then the systems now being built out of several language-model agents are architectures with the control distributed across processes, and the evidence about whether that helps sits in multi-agent intelligence systems.

Frontier Three dependencies run outward from disciplines that are not cognitive science at all. Indexing and retrieval, per the bottleneck above, is the one that matters most and the one nobody has imported. Frontier Neuromorphic hardware is the dependency for the SPAUN route specifically, and it is poorly documented in public sources: the only quantified figure this brief could establish is Neurogrid's 65,536 neurons from 2014. Speculative Mechanistic interpretability is the third, and it runs both ways — it depends on the symbolic tradition for a vocabulary of what to look for, and the tradition depends on it for the possibility that structures it specified might be found rather than authored.

Speculative One dependency is textual and mundane. The field's own consensus document was unreadable for this brief. A synthesis that cannot be read cannot be argued with, and an architectural tradition whose central recent statement sits behind a publisher's gate is competing for attention against a literature that posts everything to a preprint server the day it is written. Frontier That is not a scientific dependency, but it determines how much of this topic anyone outside it can evaluate.

Established X1 — measure the match-cost curve, then swap the index. Take Soar 9.6.5, generate rule sets spanning several orders of magnitude in size, and publish match cost per decision cycle as a function of rule count, with working-memory size held fixed and then varied. Then reimplement the match over a modern index — the same structures a database query planner or a vector store would use — and publish the curve again. Speculative The result is one plot with two lines, and it decides whether the constraint that shaped forty years of this field survives contemporary data structures. Established Cost: one competent engineer, a few months, no new science, no new hardware, no proprietary model access. Frontier It has not been published, and a negative result — the curve is still bad — would be nearly as valuable as a positive one, because it converts folklore into a measured constraint.

Established X2 — run the architectures on Psych-101. Psych-101 is exactly the dataset that makes a like-for-like comparison possible: 160 experiments, over 60,000 participants, over 10,000,000 choices, with a fitted model already benchmarked on it. Established Take ACT-R and Soar — both freely available under permissive licences — fit them to the same held-out split, and report predictive accuracy alongside free parameter counts. Frontier No architecture has been run on it, and the comparison would settle, on the field's own criterion, whether hand-built theories buy their explanatory claim at an acceptable price in accuracy.

Frontier X3 — chunking at data scale, with a language model as the content source. Soar learns rules from resolved impasses; the historical limit was that resolving an impasse required knowledge the system did not have. A language model supplies exactly that. Measure rules acquired per hour of unsupervised experience, the fraction surviving validation, and the fraction a domain expert can read and endorse. Frontier X4 — measure the interface. In any architecture with a neural component, report the repair rate: how often the component emits state the symbolic controller cannot act on without correction. CoALA specifies the architecture; no source read here measures the seam. Frontier X5 — impasse detection as a benchmark. Build a labelled set of situations in which a system lacks the knowledge to proceed, and report detection rate. Soar has the mechanism and CLARION has the metacognitive subsystem; language-model agents have neither, and knowing when you do not know is the capability the current paradigm most conspicuously lacks. Speculative X6 — parameters per phenomenon. Report, for every architecture and every fitted model, the ratio of free parameters to distinct behavioural phenomena accounted for. It is the only measurement that makes the neat position falsifiable, and it has essentially never been taken.

Established The engineering that matters here is storage and retrieval, not modelling. An architecture that acquires knowledge continuously needs a procedural store whose match is incremental and sublinear in rule count, an episodic store that can be written at the rate experience arrives and queried by partial content, and a semantic store with activation-based retrieval that does not degrade as it fills. Frontier Soar already has all three as components; what none of them has is a published scaling curve. Speculative The practical build is unglamorous: instrument the match, expose per-cycle timing as telemetry, and treat the memory subsystems as a database problem with a workload characterisation — which is how every other field with this problem solved it.

Frontier A hybrid inherits the hard parts of both components. Symbolic control needs discrete, reliable state; language-model output is neither, so the interface needs a validation and repair layer with a measured error rate, and the architecture's timing commitments have to survive a component whose latency varies by orders of magnitude between calls. Speculative Nobody has shown that a timing-fidelity claim survives that seam, and it may not, in which case the hybrid buys control at the cost of the architecture's psychological validity.

Frontier The neural route is a hardware bet with poor public documentation. SPAUN's 2.5 million neurons against Neurogrid's 65,536 is the scale gap this brief can defend; the newer large neuromorphic systems are absent from the sources read here and their neuron counts are not asserted. Speculative What SPAUN would need to be more than a proof of principle is not more neurons but learning inside the architecture and tasks that are not pre-specified — and the second is a scientific requirement, not an engineering one. Handwave A model that consumes hours of compute to produce seconds of behaviour is a demonstration about representation, and it should be read as one.

Frontier The nearest neighbour is a field the tradition has never talked to. Database indexing, query planning and approximate nearest-neighbour search solve, at industrial scale, the exact problem that bounds production systems: find the stored items consistent with a changing state, fast, as the store grows. Speculative That the utility problem was never handed to that community is the most surprising fact in this topic, and the adjacency with the highest expected value.

Frontier Retrieval-augmented generation is the same problem under a new name. Agent memory systems that retrieve from a growing store before acting are production-system match with a softer matcher, and they hit the same curve: bigger store, worse retrieval, so keep the store small or the search shallow. Speculative The architecture literature has thirty years of results about what happens next, and the agent literature is not reading them — the concrete form of the forgetting question.

Established Two adjacencies are theoretical rather than technical. Distributed cognition claims a cognitive system need not be a single organism — a navigation team is literally the cognitive unit — which turns institutions into architectures and makes the memory-action-decision decomposition applicable to groups, the same hinge that runs into collective intelligence. Frontier And machine consciousness inherits Newell's item eight: self-awareness was a constraint on unified theories, CLARION implements a metacognitive subsystem, and the question of whether functional metacognition is the whole of the phenomenon or a proxy for something else is exactly the dispute that topic runs on.

Frontier Two applied adjacencies carry evidence rather than theory. Intelligent tutoring is where an architecture met an outside evaluator, with the results given above, and it connects directly to future education systems. Speculative And memory engineering is the reverse direction: episodic memory as an engineering component was specified here first, and it is now being rebuilt for agents by people who mostly do not cite the specification.

Established This is a small field with unusually poor public accounting of itself. The research group behind Soar exists and is named — the Center for Integrated Cognition, under Laird and Wray — and a commercial spinout, SoarTech, maintains a Java implementation. Established Beyond that, the research for this brief could establish no funding figure, no roadmap, no headcount and no grant for either flagship architecture, because the pages that would carry them were behind access gates. Frontier “Is anyone funding this?” is the operative question about a research programme accused of abandonment, and this brief cannot answer it. That is a defect in the record, and the fix is a page on a website.

Frontier The tradition has no benchmark culture, and that is why it cannot defend itself. The field counts publications — more than 700 for ACT-R, over 900 practical projects across the surveyed architectures — where the competing paradigm counts scores on shared tasks. Established Publication counts are not comparable across paradigms; benchmark scores at least pretend to be. Speculative A community that had maintained a shared evaluation suite for thirty years could answer the abandonment charge with a curve instead of an anecdote, and the absence of that curve is a choice made repeatedly over decades, not an accident.

Speculative The citation asymmetry between the two literatures is the institutional finding. CoALA is the exception that shows the rule: it connects language-agent research to classical cognitive science explicitly, and it is notable for doing so. Frontier The broader agent literature reinvents working memory, procedural versus declarative knowledge, subgoaling, reflection and episodic recall largely without attribution, and the reinvention is often worse than the original because the original had a control regime and a decision cycle rather than a prompt. Handwave Whether that is anyone's fault matters less than whether it is fixable. One well-cited survey mapping agent-framework components onto the architectural primitives they duplicate would cost a person a summer and would be immediately useful to builders who do not know that thirty years of results about their design problem exist.

Established The tradition's distinguishing property is the one that accountability regimes ask for. A production rule can be read, cited in an explanation and edited; a learned rule from chunking is, in principle, inspectable in the same way. Frontier That makes cognitive architectures the one route to capable behaviour whose reasons are stateable in advance rather than reconstructed afterwards, which is why the inspectability conjunct in the workback target is not a stylistic preference — it is the property that would make such a system usable where a decision has to be justified. Speculative It is also the property automatic acquisition threatens: rules learned at scale may be as unreadable as weights, and no source read here reports what fraction of chunked rules an expert can endorse.

Established Two deployment records raise different questions. The tradition's largest fielded system was military: synthetic tactical air agents in a joint battlespace exercise, which is a use of a theory of human cognition to simulate adversaries and allies at scale. Established Its largest civilian system is a school tutor whose evidence base includes a statistically significant negative result in one of six methodologically adequate studies. Frontier Neither is scandalous, and both are worth stating plainly: a field mostly discussed in the abstract has two concrete records, and they are a war game and a maths curriculum.

Speculative A predictive model of human choice is an instrument with a second use. Centaur is trained on ten million human choices to predict what people will do in situations they have not seen. As science that is a challenge to architectural theories; as an artefact it is a general-purpose model of human decision-making under experimental conditions. Frontier The architectures never had this property, because a theory you can read is a theory whose limits you can see. Handwave The ethical asymmetry between an explanatory model and a predictive one is under-discussed in this literature and probably belongs in the next revision of it.

Speculative The programme's claim, if it were true, would be one of the largest claims in science. A unified theory of cognition asserts that minds have a fixed structure — not this mind or that species, but minds — and that the structure is small enough to write down. Speculative That is a claim about every intelligence a civilisation might build, meet or become. Handwave Newell's list already made it a claim about biology too: neural realisability plus evolutionary emergence means the structure has to be the kind of thing selection can produce.

Frontier The deflationary reading is that the tradition's legacy is a vocabulary, and vocabularies are how paradigms survive defeat. Working memory, procedural versus declarative, chunking, impasses, decision cycles, buffers, episodic memory: every one is now standard in agent engineering, usually with no attribution. Speculative If that is the whole legacy, the architectures were a research programme that produced concepts rather than systems — not a failure, but a different thing from what was promised.

Speculative The civilisational stake is inspectability. There is a version of this topic in which nothing turns on whether Soar or ACT-R was right, and everything turns on whether a capable system can be built whose reasons are written down before it acts. A society running its institutions on systems whose reasons cannot be stated has a governance problem that no amount of capability fixes, and the architectural tradition is the only serious attempt at capable systems whose reasons are readable by construction. Handwave If the utility problem turns out to be an artefact of old data structures, the choice between capable-and-opaque and readable-and-limited was never forced, and a civilisation spent a generation on one horn of a false dilemma.

These horizons track the programme rather than the hardware: when a fixed-structure theory of cognition could be shown to scale, and when the abandonment question would be settled rather than argued. Each horizon is written so that a null result counts — a field this small can only be moved by experiments whose failure is publishable.

• 10 yr: Frontier The two cheap experiments are run or the field has visibly declined to run them: the match-cost curve measured against a modern index, and an architecture fitted to Psych-101 with free parameters counted on both sides. Agent frameworks converge further on the architectural primitives by construction — persistent episodic stores, explicit control loops, self-monitoring — and the citation gap either closes through one good survey or hardens into two literatures sharing a vocabulary and no results.
• 25 yr: Speculative Automatic knowledge acquisition into architectural form is demonstrated at some scale, with an inspectability figure attached — the fraction of learned rules a domain expert endorses. If that fraction is high, the tradition has its answer to the knowledge bottleneck; if it is low, inspectability was never separable from hand-authoring and the programme's distinctive claim collapses.
• 50 yr: Speculative Monotonic improvement over months without an author editing the knowledge base is either demonstrated or shown to be structurally impossible for a system that stays readable. This is the point at which the abandonment question is answered rather than argued, and it is answerable in either direction.
• 100 / 250+ yr: Handwave Either a small set of mechanisms is known to account for cognition with fewer free parameters than the behaviour it explains — the neat position vindicated, Newell's nine constraints met including the two he was ignored about — or the scruffy answer is settled: there is no single principle, intelligence is what a large enough fit of a large enough world looks like, and the architectural vocabulary survives as engineering convention long after the theory it came from stopped being believed.

• Depends on Waits on two results this map produces elsewhere. Every claim in this topic is a comparison — architecture against fitted model, scaffold against end-to-end system, an architecture against its own earlier version — and comparisons need the protocol discipline and, specifically, the free-parameter accounting set out in intelligence measurement: a hand-built theory beating a fit, or losing to one, means nothing until both sides' parameters are counted the same way and the evaluation split is named. The dependency is not decorative. The single experiment this brief most wants run, fitting Soar and ACT-R to Psych-101 against Centaur, is uninterpretable without it. It also waits on multi-agent intelligence systems, because a control layer distributed across several agents is an architecture with its decision cycle spread over processes: whether that composition beats one model doing the whole job is the same question as whether a control layer earns its overhead, and the compute-matched evidence is being gathered there rather than here.
• Enables Supplies the control vocabulary that other topics keep rebuilding. Episodic memory that records state automatically, impasse-driven subgoaling, metacognitive monitoring and a decision cycle with graduated commitment are architectural primitives with thirty years of results attached, and they are simultaneously the open problems of memory engineering and of agent construction generally. Any account of artificial general intelligence that expects one fixed structure sufficient for all tasks is making this topic's claim, whether or not it uses the word architecture — and inherits this topic's binding constraint along with it, because a system that accumulates knowledge indefinitely has to retrieve from it indefinitely.
• Adjacent Runs alongside distributed cognition and collective intelligence, where the unit of cognition is a group and the same memory-action-decision decomposition applies; alongside machine consciousness, which inherits the self-awareness constraint from Newell's own list; alongside future education systems, where the tradition's one commercial artefact has an outside evaluation record; and alongside database indexing and retrieval, which is not a topic on this map and is where the binding constraint of this one has been solved for other purposes.

Established “Cognitive architectures are a historical curiosity.” The census says 49 of 84 surveyed architectures are actively developed; Soar shipped 9.6.5 with a new chunking operator, a CMake migration and SWIG bindings for five languages; ACT-R stands at 7.21.6; ACT-R models appear in more than 700 publications. Frontier The field is small, not extinct — and the two flagships are in different states, which is a distinction worth preserving rather than averaging away.

Established “SPAUN simulated a brain.” It simulated 2.5 million neurons in structures resembling prefrontal cortex, basal ganglia and thalamus, performing numeric microtasks. Established The two independent sources read here enumerate four capability areas and give no task count, no speed, no hardware and no successor. Frontier The widely circulated figure of eight tasks is not printed here, and anyone quoting it owes a source this brief could not obtain. Speculative For scale, 2.5 million neurons is a small fraction of one per cent of a human cortex on any standard estimate; the model's importance was always representational rather than dimensional.

Established “Symbolic AI failed and neural networks won.” Symbolic AI dominated from the mid-1950s to the mid-1990s; the second AI winter ran 1988 to 2011; neural networks were considered hopeless until 2012, when GPUs arrived. Established Both traditions spent decades in disgrace, and the current frontier is explicitly rebuilding symbolic control structures on top of neural components. Frontier One approach is ascendant on a timescale shorter than either tradition's dormancy.

Established “Symbolic and connectionist are opposed.” Smolensky's tensor product variable binding lets connectionist systems manipulate symbolic structures; CLARION runs both levels inside one architecture; and the reference account of connectionism now states the two “need not be at odds,” suggesting the live division is computationalism versus dynamical systems. Speculative What survives is a dispute about learnability rather than expressibility: not whether a distributed system can represent structure, but whether structured representations can be acquired by gradient descent from realistic data. That is open, and large language models are the running experiment.

Established “Newell and Simon won the 1976 Turing Award.” The award was 1975. The lecture is dated 1976 because that is when it was published. It is a small, common, checkable error and it appears constantly.

Established “Cognitive Tutor is proven to work.” The 2013 Institute of Education Sciences review examined 27 studies, of which 6 met rigorous standards: 5 positive, 1 statistically significantly negative. The 2009 Algebra I review found potentially positive effects in 1 qualifying study out of 14. Frontier That is a modest evidence base with a documented negative result inside it — and still better evidence than most educational technology has.

Frontier Enthusiast-side error: “LLM agent frameworks are a new kind of cognitive architecture.” On the field's own account they are a re-derivation. CoALA's “modular memory components, a structured action space ... and a generalized decision-making process” is the skeleton of Soar and ACT-R, and the paper says so, connecting language agents explicitly to classical cognitive science and symbolic AI. Speculative Naming the lineage is more useful than claiming novelty: if the decomposition was forced, then the tradition's results about what happens next — what learned rules do to retrieval, what a decision cycle needs in order to commit — are results about the new systems too.

Frontier “Neuro-symbolic AI is delivering.” It has a six-category taxonomy and a dozen named systems, and the reference account read for this brief contains no quantitative benchmark results at all. Established No neuro-symbolic benchmark number appears in this brief. Speculative A programme with a rich taxonomy and no scoreboard is not necessarily failing, but it cannot be said to be succeeding either.

Speculative “Chunking is how Soar learns, therefore Soar scales.” Chunking is the mechanism and the utility problem is the reason it is not sufficient: match cost grows with the rule set, and learning more rules can make a production system slower. Frontier No source read here says the problem is solved, and none reports a measurement against a current release. Handwave The most interesting thing in this brief is that the unsolved problem is an indexing problem, that indexing has been transformed since the 1980s, and that the check is a few months of work nobody has published.

Established “Cognitive architectures don't connect to neuroscience.” ACT-R has been extended to predict BOLD response shapes in motor cortex, prefrontal cortex, anterior cingulate cortex and basal ganglia by mapping buffers onto regions, and SPAUN is built from 2.5 million spiking neurons in region-organised structure. Frontier This is the tradition's most under-reported strength, and the thing lost if the programme were replaced wholesale by fitted models: a prediction about brain imaging is a different order of commitment from a prediction about choices.

Frontier Skeptic-side error: “a fitted model beating architectures settles it.” Centaur's reported result — better prediction of held-out participants than existing cognitive models, on Psych-101 — is a serious challenge on the criterion the architectures chose. Speculative It is not yet an answer, for two statable reasons: no architecture has been run on that dataset, so the comparison has not been made head to head with parameters counted; and prediction was never the whole architectural claim, which was that the structure is a theory. Frontier Whether that reply is principled or defensive is the live question, and X2 would move it.

Established Declared gap: the Common Model of Cognition. The Laird-Lebiere-Rosenbloom synthesis is the field's own consensus attempt, and this brief could obtain no account of its content — the encyclopedia article was cache-blocked across repeated attempts and only a bibliographic record for a 2026 MIT Press chapter was verified. Established No number, component count or structural claim from it appears anywhere above. Frontier Treat any summary of the Common Model, including a confident one, as owing a source — and notice that the tradition's consensus document is harder to obtain than the preprints challenging it.

Handwave The exotic claim, stated at full strength: the programme was not refuted, it was outspent. The version worth taking seriously runs like this. The architectures were never given compute — the entire tradition ran on academic budgets, and the 2012 turn happened because GPUs arrived, an experiment nobody ever ran for an architecture. They solved problems the current paradigm has re-encountered and not solved: automatic episodic recording, impasse-driven subgoaling, metacognitive monitoring. Their binding constraint was authorship cost, and a language model is exactly a fix for authorship cost. Their fatal technical limit was an indexing problem from the era of expensive memory. Speculative What would have to be true for this to be right is specific: an architecture must be shown to benefit from scale, which has never been tested; learned rules must stay inspectable, which has never been measured; and retrieval must not degrade, which is the unrun experiment. Established The mainstream reply is equally specific and should be stated plainly: the tradition had forty years, its largest deployment is from 1997, its psychological validation is a list of laboratory microtasks, its knowledge bottleneck was never solved from the inside, and a field that could not produce a shared benchmark in three decades does not get to attribute its position to funding. Frontier Both accounts are consistent with everything this brief could verify. That is an honest terminal position, and it is also an unusual one, because the experiment that would break the tie is cheap, specific, and has not been run.