Concepts · Humans and machines
SOAR chunking creates new productions from resolved impasses so similar situations can be handled faster later.
Chunking in SOAR
How does SOAR learn from getting stuck?
One answer is chunking.
In ACT-R, a chunk is a unit of declarative memory. In SOAR, the word is used differently. SOAR chunking creates a new production from a solved problem.
That difference matters.
A SOAR chunk is not just a stored fact. It is learned procedural knowledge. It helps the system avoid doing the same deliberate problem solving again.
From subproblem to rule
SOAR reaches an impasse when it cannot move forward. It creates a substate. The system works inside that substate until it finds a result that lets the higher-level problem continue.
When that happens, SOAR can summarize what mattered.
The learned chunk says, in effect:
When a similar situation appears again,
produce the useful result directly.That is why chunking is powerful. It turns slow problem solving into faster future behavior.
A simple example
Suppose an agent keeps facing the same support ambiguity.
The customer is in a high-trust tier. The refund is below a safe limit. The product was purchased recently. The policy allows an automatic refund.
The first time, the agent may need to check several things. It may create a subtask, inspect policy, compare options, and resolve the impasse.
After that, a learned rule could help:
If the customer is high-trust
and the refund is below the safe limit
and the purchase is recent,
prefer the automatic refund operator.That is not official SOAR syntax. It shows the shape of the lesson.
Why AI builders should care
AI systems often repeat work.
They re-read the same policy. They re-derive the same tool sequence. They recover from the same error in the same slow way.
SOAR chunking gives a useful design question:
What should the system learn from the way this problem was solved?
The answer should not be "save the whole transcript." A transcript may help, but it is noisy. The useful lesson is often smaller: the condition that mattered and the action or preference that resolved the problem.
The caveat
Learned rules need review.
If the system learns from a bad resolution, it can become faster at doing the wrong thing. If the situation changes, an old chunk can mislead future behavior.
So the AI lesson is not "auto-learn every rule."
The lesson is: when a repeated problem is solved well, preserve the useful pattern in a form that can be inspected, weakened, or removed.
The takeaway
SOAR chunking turns solved problem solving into future procedural knowledge.
For AI agents, it points to a better kind of memory: not only saving what happened, but learning which pattern helped the system move forward.
Sources
Go back to Impasses and Subgoals in SOAR, continue with Semantic and Episodic Memory in SOAR, or return to the Concepts hub.