Among his other contributions to artificial intelligence and cognitive science, Allen Newell defined a framework in which to interrelate cognitive architectures and define intelligence.
A fundamental notion is that of the knowledge system, which is an agent in an environment that uses the knowledge it has internally represented to achieve its goals. Represented knowledge is stored in a stable but malleable medium that may be transformed in regular ways. There is a mapping between transformations in the environment and transformations in the medium. One way to support knowledge manipulation is the use symbols, wbich are distinguishable patterns that may be mechanically manipulated, and be used to index additional information.
An architecuture is the fixed structure that realizes a symbol system. Because, for a give system, different aspects may be considered "fixed" or "variable", the concept of "architectural" varies with time scale. In implementing a symbol system, an architecture provides a clean, discretized, symbolic world for the agents on top.
Newell defined intelligence in terms of how well a system approximated a knowledge-level system. What makes knowledge-level systems intelligent is their ability to use all information they have to achieve their goals. To the extent that a system does this, it is intelligent.
A system uses its knowledge to achieve its goals. If necessary knowledge it not immediately on hand, it sets up a subservient goal to look for that knowledge. The subservient goal and its circumstance are a problem space, and the act of finding the goal is that of search. In SOAR the act of instantiating a new problem space to get missing information is called universal subgoaling.
To achieve a given level of competence, a system may try to store a lot of search results, or it may try to be fast enough to generate and search problem spaces dynamically. Humans do the first, while some of the best AI programs to the second. Newell states that a cost tradeoff between them exists, and that some balance gives the best performance for cost.