Lesson 12 · Solution · SOAR II: chunking as learning

Solution: Never Solve the Same Impasse Twice

Correct: ACT-R maintains a separate declarative memory with graded, activation-based retrieval; SOAR compiles experience directly into new all-or-nothing production rules, with no separate declarative store in the classic architecture.

Both other mechanism-comparison options are wrong on the facts: SOAR clearly does learn (chunking is its learning mechanism), so “only SOAR learns” is false, and chunking and base-level activation are not the same thing — one produces a new discrete rule, the other continuously reweights an existing fact’s retrievability. “ACT-R only models procedural skill” is backwards, if anything — ACT-R is unusual precisely for modeling both declarative and procedural memory as separate, interacting systems.

Why this is a real design bet, not a footnote. Committing to “only production rules, nothing else” (SOAR) versus “facts and rules are genuinely different kinds of memory” (ACT-R) isn’t a style choice — it’s a claim about the mind with testable consequences. ACT-R’s bet lets it predict things SOAR’s architecture doesn’t naturally produce: partial, graded recall (a name on the tip of your tongue — retrievable, but weakly, exactly what lesson 10’s negative-but-not-infinitely-negative activation captured), and forgetting that’s continuous rather than all-or-nothing. SOAR’s bet predicts something different: skills that are either fully automatic (chunked, instant) or fully effortful (not yet chunked, requires full deliberate subgoal search) — a much sharper line than ACT-R’s smooth activation curve, and a genuinely different empirical claim about how skill acquisition should feel and time out as it develops.

Neither architecture is simply “more right.” Chunking’s crisp automatization story fits skill-acquisition data (a novice driver’s effortful gear-shifting becoming instant reflex) better than a smoothly graded activation curve would predict for that kind of case. ACT-R’s graded activation fits partial/uncertain recall (the tip-of-the-tongue state, or confidently misremembering a barely-active fact) that a strictly binary “chunked or not” story struggles to represent at all. Real cognition seems to need both flavors of forgetting-and-remembering somewhere — which is a live, unresolved argument between the two research communities, not settled by either architecture alone.

In a modern LLM agent harness, chunking rhymes with caching a resolved sub-task’s answer as a reusable fixed procedure (a memoized tool call, a saved successful plan) so the next occurrence skips straight to the cached result — while a separate, editable memory store the agent can also consult, reason over, and update mid-task (not baked into fixed behavior) is closer to ACT-R’s declarative side. Harnesses that only cache-and-replay lean SOAR; harnesses with a queryable memory store lean ACT-R — most real systems, tellingly, end up wanting both.

Where this goes: last lesson in this stage steps back from either architecture’s internals to ask the harder question — how well do either of these detailed mechanisms actually predict real human timing and error data, and where do they fall short?

How was this one? Any answer marks it complete and moves on — your rating shapes future lessons.