SOAR Cognitive Architecture: States, Operators, Impasses

SOAR is a cognitive architecture for goal-directed problem solving.

Developed by Allen Newell, John E. Laird, and Paul S. Rosenbloom, and maintained by the Soar research group at the University of Michigan, it models intelligent behavior as a cycle: a system inspects a state, selects an operator that can change it, applies the operator, and checks whether the goal is reached.

text

state -> select operator -> apply operator -> new state -> repeat

This guide is Achiral's concept reference, not the official Soar project site. The official manual and source code live at soar.eecs.umich.edu.

SOAR components at a glance

Component	What it is
State	The current problem-solving situation, represented as working memory elements
Operator	A candidate action that can change the state
Goal	The target outcome the system works toward
Impasse	A stuck point when no operator can be selected or applied
Substate	A new problem-solving context opened to resolve an impasse
Chunking	Learning productions from the results of substate problem solving
Working memory	Active context the architecture reads and writes during the decision cycle
Semantic memory	Long-term general knowledge, accessible through retrieval
Episodic memory	The agent's recorded stream of experience

The SOAR decision cycle

SOAR works by repeatedly selecting and applying operators.

A state is the current situation. An operator is a possible action that can change the state. A goal is the outcome the system is trying to reach.

Imagine an assistant helping a user with a support issue. The current state might include the user, the task, the last tool result, and the open goal. Several operators may fit:

ask for more information
check account status
retry the tool
escalate to a human

SOAR represents those as candidate operators, runs a decision procedure, chooses one, applies it, and updates the state. The loop continues until the goal is reached or an impasse is detected.

What happens at an impasse

When SOAR cannot select or apply an operator, because none are proposed, preferences conflict, or multiple operators tie, it creates a substate.

The substate is a new, smaller problem: resolve the stuck point. SOAR works inside it until it finds a result that lets the higher-level state continue.

This is not a crash. It is a deliberate way to handle incomplete knowledge. When substate problem solving creates a result, chunking can summarize the relevant reasoning in a new production.

SOAR vs ACT-R

Both are cognitive architectures, and both use the word "chunk." Keep the split clear:

Topic	Difference
Design emphasis	SOAR foregrounds operator selection and impasses; ACT-R foregrounds buffers, activation, and declarative retrieval
"Chunk"	In SOAR, a chunk is a learned production; in ACT-R, it is a unit of declarative memory
Control loop	SOAR proposes, selects, and applies operators; ACT-R matches and fires productions over buffer contents
AI design lens	SOAR helps explain how a next step is chosen; ACT-R helps explain how memory becomes available and guides action

If your question is about memory activation and retrieval, start with ACT-R Memory Architecture. If your question is about decision cycles and agent problem solving, stay here.

Why SOAR matters for AI agents

AI builders often describe an agent as a language model plus tools.

SOAR asks a sharper question: how is the next step chosen?

The answer cannot be "the prompt decides" indefinitely. A useful agent needs a represented state, candidate actions, rules for choosing among them, and a way to learn when it gets stuck.

SOAR gives names and structure to those parts. It does not dictate permissions, privacy, or product boundaries, but it clarifies where the decision-making problem actually lives.

What to remember

SOAR treats intelligent behavior as a repeating cycle of state inspection, operator selection, and action. When that cycle stalls, it opens a subproblem. When the subproblem produces a result, chunking can turn the relevant reasoning into a faster future rule.

For AI memory systems, the implication is direct: memory should not only record the past. It should help the system choose what to do next.

Sources

Next, continue with Working Memory in SOAR, Operators in SOAR, or return to the Concepts hub.

What Is SOAR Cognitive Architecture?