A cognitive architecture is the part of an intelligent system that does not change when it learns. It defines how memory is organised, how attention is allocated, how goals are represented, and how perception, reasoning and action connect. Knowledge goes in. The architecture decides what can be done with it.
Classic examples
ACT-R models how declarative and procedural memory interact. SOAR introduced problem spaces and chunking. LIDA borrows from global workspace theory. Each was an attempt to specify the machinery of cognition independently of any particular task.
Why it matters for AGI
Adding capability by adding modules produces a system that grows sideways without becoming coherent. An architecture is the claim that the parts fit together in a principled way, and that new abilities can be added without breaking the ones already there.
The modern version
Today's agent stacks, with memory stores, planners, tool routers and evaluators, are cognitive architectures in all but name. Most are assembled ad hoc. The ones that work tend to have a clear theory of what each part is for.