Cognitive Architecture
A cognitive architecture is a fixed, domain-independent specification of the memories, representations, processes, and learning mechanisms that together produce intelligent behavior, proposed either as a theory of how the human mind is organized or as a blueprint for building general artificial agents. The architecture is the part that stays constant across tasks, in the way a computer's instruction set stays constant across programs; what changes from task to task is the knowledge loaded into it. The field treats intelligence as a systems-level property that emerges from the interaction of memory, perception, decision-making, and learning components rather than from any single clever algorithm, which sets it apart from approaches that optimize one narrow function at a time.
The intellectual founding statement is Allen Newell's 1990 book Unified Theories of Cognition, which argued that psychology had accumulated a mountain of isolated experimental effects and needed a single integrated theory able to account for many of them at once. Newell's 1973 essay "You can't play 20 questions with nature and win" made the earlier version of the complaint: carving behavior into binary experimental contrasts never adds up to a theory of the whole mind. His answer was to build working computational systems whose every mechanism is committed in advance and then to test them across many tasks. The lineage runs back through Newell and Herbert Simon's General Problem Solver and their production-system models of human problem solving, where knowledge is encoded as condition-action rules that fire when the current state of working memory matches their conditions.
Soar, begun by John Laird, Allen Newell, and Paul Rosenbloom in the early 1980s and described in their 1987 paper "Soar: An architecture for general intelligence," represents all problem solving as search through problem spaces and uses a single learning mechanism called chunking. Its decision cycle gathers every production that matches the current state, and when the available knowledge does not settle which operator to apply, an impasse arises and Soar automatically creates a subgoal to resolve it. When the subgoal succeeds, chunking compiles the reasoning that happened inside it into a new production, so the same situation is handled in one step next time. Later versions added reinforcement learning, semantic memory, episodic memory, and mental imagery to the original procedural core, and Soar has been used for real-time agents in simulated military training environments and in robotics research.
ACT-R, developed by John Anderson at Carnegie Mellon from his earlier ACT* theory (The Architecture of Cognition, 1983) and The Adaptive Character of Thought (1990), is the architecture most tightly bound to human experimental data. It divides the mind into modules for vision, motor control, declarative memory, and goals, each exposing a small buffer, with a central procedural system selecting one production at a time to fire. Declarative knowledge is stored as chunks whose retrieval depends on a numeric activation reflecting recency, frequency of use, and similarity to the retrieval cue, which lets the architecture predict reaction times, error patterns, and forgetting curves. Anderson and colleagues' 2004 paper "An integrated theory of the mind" argued that its modules could even be mapped to specific brain regions, tested against functional MRI data. The neighbors in this family include EPIC, which emphasizes perceptual-motor timing, and CLARION, which explicitly separates implicit and explicit knowledge.
The two main architectures embody different goals: Soar is principally an engineering and generality-focused system, while ACT-R is principally a model of human cognition constrained by timing and error data. A frequent misconception is that a cognitive architecture is simply a program that solves problems; strictly, it is a theory whose fixed mechanisms make commitments that can turn out to be wrong, and its worth is judged by how many tasks it explains without per-task tuning. A related pitfall is that the knowledge-engineering burden is large: hand-writing productions and chunks for a rich domain is slow, which is a major reason architectures of this kind were overshadowed by statistical learning for decades.
The idea has returned in the era of large language models, where agent frameworks bolt a model onto memory stores, tool interfaces, and a planning loop. The 2023 paper "Cognitive Architectures for Language Agents" by Theodore Sumers, Shunyu Yao, Karthik Narasimhan, and Thomas Griffiths explicitly borrowed the vocabulary of Soar and ACT-R, describing working memory, long-term memory, action spaces, and decision procedures around a language model. The classical systems supply what a bare model lacks: persistent memory, explicit control flow, and learning that accumulates across episodes. Whether a single learned network or a modular architecture is the better route to general intelligence remains an open design question, and neurosymbolic hybrids are one active compromise.
Read cognitive architecture as a theory of the mind written as running software: a fixed machinery of memories, a decision cycle, and learning mechanisms that stays constant across tasks, so that any behavior it produces must be explained by knowledge loaded into that machinery rather than by tuning the machinery for each new problem.