Cognitive architectures: integrating capabilities does not prove general intelligence

In September 1987, John E. Laird, Allen Newell and Paul S. Rosenbloom published SOAR: An Architecture for General Intelligence. The title declared a research programme: build a computational organisation capable of supporting many cognitive tasks. It did not announce that artificial general intelligence, or AGI, had been achieved.

That distinction remains the best tool for reading a cognitive architecture. An architecture specifies relatively stable components—memories, representations, learning, decision and action—and rules for coordinating them. Knowledge for a particular task is added later. Integration is a property of the design; generality is an empirical claim that requires evidence of breadth and transfer.

The transferable skill is to separate five levels that are often merged: architecture, added program or knowledge, environment data, evaluated task and measured result. When a headline jumps from any of the first four to “general intelligence,” this chain identifies exactly which evidence is missing.

What a cognitive architecture is

A cognitive architecture is an implementable hypothesis about the mechanisms that produce intelligent behaviour. It may specify how information enters, what remains active, how memories are retrieved, how candidate actions compete and how experience changes later decisions. Its ambition is to reuse that framework across tasks rather than design an incompatible system from scratch for each one.

The term does not require copying every biological detail. SOAR organises symbolic states, operators and rules. ACT-R officially describes itself as a theory for simulating and understanding human cognition. Other approaches combine symbolic modules with neural learning. They share an integration question but can pursue different goals: build an effective agent, explain human response times or explore a computational theory of mind.

“Cognitive architecture” is therefore not a synonym for AGI. A spreadsheet has coordinated modules and memory, but it is not general. The cognitive label says which functions a design attempts to organise and which phenomena it seeks to explain. Generality must appear in behaviour: new tasks, transfer, adaptation and resilience to context changes.

SOAR: architecture is not program

SOAR's official manual states a central hypothesis: deliberate goal-directed behaviour can be cast as selecting and applying operators to a state. A state represents the current problem situation; an operator proposes a transformation; a goal specifies the desired outcome. The architecture runs cycles that propose, compare, select and apply operators.

Working memory holds the present situation, including sensory data, inferences, active goals and operators. Production memory stores procedural knowledge as rules resembling “if this, propose or apply that.” Preference memory helps choose among operators. When available knowledge cannot make progress, SOAR creates a substate to resolve that blockage, or impasse.

The limitation appears in the same documentation: the architecture cannot solve any problem without added procedural knowledge. SOAR is the shared mechanism; a SOAR program supplies the rules and representations needed for a task or family of tasks. Crediting the framework with every capability in a demonstration would be like crediting a programming language with the knowledge contained in a program written in it.

One learning mechanism, chunking, summarises the processing that resolved an impasse and creates a new production. Next time, the system may bypass part of the intermediate reasoning. This demonstrates a route for procedural learning within the architecture. It does not guarantee transfer to every domain or harmless generalisation; both must be tested.

Three questions that prevent category errors

The first question is whether a project proposes general mechanisms or optimises a particular function. A neural network can learn millions of parameters for a task and form part of a larger architecture, but size alone does not make it an integrated theory of memory, goals and action. Conversely, a symbolic design does not gain robust perception by including a box labelled “vision.”

The second question is whether a system produces behaviour or simulates a natural phenomenon. A cognitive architecture may predict how a person remembers or decides; a biological reconstruction may reproduce electrical properties of tissue. Both can inform the study of mind, but faithfully simulating one part does not give the system every function of the complete organism.

The third question is whether evidence shows depth in one domain or transfer across domains. An agent can outperform human experts under closed rules and have no mechanism for acting outside them. Specialised excellence is strong evidence about that task, not weak evidence about all the others.

AlphaGo: specialised integration

The AlphaGo paper published on January 27, 2016 describes a precise combination: a policy network selected moves, a value network evaluated positions and Monte Carlo tree search explored continuations. The networks learned first from expert games and then through reinforcement learning from self-play. The system defeated European champion Fan Hui five games to zero.

AlphaGo integrates board perception, evaluation, learning and planning through search. In that sense it presents a sophisticated decision architecture. Every component, however, shares one interface: legal Go positions and moves, a final win-or-loss reward and stable rules. The work does not show that the same system can learn medicine, negotiate a requirement or use everyday language without redesign and domain-specific data.

The lesson is methodological. A generality claim cannot rest on counting modules or noting that an agent learns. It must measure how much of the mechanism, knowledge and representation survives a task change. If everything but the label must be rebuilt, transfer is small even when original performance is extraordinary.

Blue Brain: biological simulation is not a general agent

The Blue Brain Project addressed another question. Its official description sets a goal of building biologically detailed digital reconstructions and simulations of the mouse brain. In 2015 it presented a first reconstruction of neocortical microcircuitry. The primary Cell paper defines the object: 0.29 cubic millimetres of juvenile rat somatosensory cortex, roughly 31,000 neurons, eight million connections and 37 million synapses.

Those numbers describe resolution and scale inside a specific tissue. They do not describe a complete human brain, a mind or a goal-directed agent. The team called the reconstruction a first draft built from sparse experimental measurements and organisational principles. Its value was to reproduce and explore physiological phenomena and generate predictions that experiments could challenge.

SOAR and Blue Brain work at different explanatory levels. SOAR asks which computational organisation can coordinate problem solving. Blue Brain asks how biological structures and dynamics can be reconstructed. A functional explanation can omit cellular detail; a cellular simulation may not specify goals, task learning or action. Neither automatically substitutes for the other.

A matrix for evaluating generality

Breadth: how many different task families can the system complete under a shared protocol? Solving a hundred variations of one game adds depth, not necessarily breadth. Tasks must demand different capabilities without a prepared interface doing most of the work.

Adaptation cost: can it learn a new task from instructions or a few examples, or does it require a full engineering and training cycle? A system may appear in many domains because specialists built different programs on one architecture. That proves framework reuse, but not autonomous cross-domain learning.

Transfer and retention: does one skill reduce the effort needed to learn another, and does new learning preserve the old? The test must separate positive transfer from shared hardware and measure interference or forgetting as well.

Robustness: does behaviour survive changes in phrasing, noise, partial goals or environment? A prepared demonstration can hide rigid assumptions. Out-of-distribution and adversarial tests reveal whether the system captured a pattern or memorised its surface.

Autonomy and correction: can it detect missing information, request help and allow decisions to be stopped or reviewed? More integration creates possibilities while chaining errors together. A record of which memory, rule, observation or operator influenced an action can matter more than a generic architecture diagram.

Applications that demonstrate something specific

A useful cognitive-architecture demonstration should identify the participating components and the measured outcome. A robot following a hierarchical instruction may show interaction among perception, working memory, operator selection and learning. A behavioural simulation may compare model and human response times or errors. A tutor may represent a learner's state and adapt an action under testable rules.

Not every AI product in medicine, search or industry is a cognitive architecture. A commercial name does not reveal memories, mechanisms or control flow. Supporting that label requires system documentation, an experiment, a task and a result. Without that chain, an example shows only that an application contains AI, not that it tests a theory of cognition.

What can be claimed about AGI

An architecture aimed at generality can be a valuable engineering condition: it forces decisions about sharing knowledge, learning and acting. It is not a certificate. Evidence would have to show new-task learning, transfer, retention, combined perception and action, robustness outside a prepared environment and recognisable limits.

SOAR shows that a framework can coordinate memory, rules, operators, subproblems and procedural learning. AlphaGo shows that networks and search can be integrated for exceptional Go performance. Blue Brain shows the scientific value of reconstructing neural tissue in detail. Each achievement remains real precisely because its denominator remains attached.

Separating architecture from program, integration from transfer and simulation from behaviour provides a durable test for future claims. The decisive question is not how many components a system has, but which capabilities survive when the problem changes and how much outside work adaptation requires. That is where evidence of generality begins; before it, there is an interesting proposal rather than proof.