Cognitive Architecture

A cognitive architecture is a blueprint for a system meant to think. It specifies which components exist and how they communicate with one another. Typical components include a part for perceiving the environment, a part for remembering, and a part that decides what to do next. The term originates from research into human thought. Researchers wanted to know whether all of the mental feats of humans could be explained by a single underlying framework. Out of this question emerged computer programs that rebuild and test this framework.

A single framework instead of a thousand individual solutions

Most programs solve exactly one task. A chess program plays chess and nothing else. A cognitive architecture pursues a larger goal: it aims to use the same framework for chess, for steering a vehicle, and for solving puzzles. Only the knowledge inside changes, while the framework remains the same.

This is also useful for research on humans. If a model takes just as long as a human to complete a thinking task and makes the same errors, that is a telling clue. It may mean the assumptions about human memory are correct. Cognitive architectures are therefore tools of psychology, not just of computer science.

Currently, the topic is gaining new momentum. Large language models can generate text astonishingly well, but forget everything once the conversation ends. Anyone wanting to build an assistant out of such a model that works on a task over the course of days needs memory, goals, and planning. These are exactly the questions that cognitive architectures have addressed for decades.

Memory, rules, and the constant cycle

Almost all cognitive architectures distinguish between two kinds of memory. Working memory contains whatever is currently important, i.e., the present situation. Long-term memory contains everything that has been learned, such as facts and proven procedures. Together, both make up the state of the system at any given moment.

On top of this lies a cycle that runs continuously. First, the system looks at what is in working memory. Then it searches long-term memory for matching rules of the form “if this holds, then do that.” Often several rules fit at once, so one must be selected. The chosen rule changes working memory or triggers an action, and the cycle begins anew.

Two examples define the field. SOAR was developed starting in the 1980s and treats every problem as a search for the next sensible step. ACT-R comes from psychology and models how quickly humans recall something. A common misconception is that a cognitive architecture is itself already a finished artificial intelligence. Without knowledge filled in, it is empty, like a shelf without books.

From flight simulators to AI agents

In practice, cognitive architectures are found wherever machines are meant to convincingly imitate human behavior. Military simulations use them for opponents that react realistically. Vehicle manufacturers use them to model how a driver reacts to a sudden braking maneuver. Learning software also relies on them to identify exactly where a student is getting stuck.

In the news today, the term mainly appears in connection with AI agents. This refers to systems that autonomously carry out multiple steps in sequence, such as doing research and then making a booking. When a company states that its agent has a memory and a planning component, it is essentially describing a cognitive architecture. The difference from before: the actual thinking is often handled by a language model rather than hand-written rules.

The term should not be confused with the architecture of a neural network. That concerns layers and computational pathways within a single model. A cognitive architecture describes the level above that, namely the interplay of entire building blocks. Both can occur at the same time, but they mean different things.