Lesson 11 left a SOAR agent resolving an impasse by subgoaling — sometimes an expensive search through a whole new problem space just to break one tie. SOAR’s second major mechanism, chunking (an unrelated reuse of Miller’s word from Stage 0 — pure terminology collision, not the same idea), is what stops that expensive search from ever repeating.
When a subgoal resolves, SOAR doesn’t just apply the result and discard the reasoning that produced it. It automatically compiles the trace — the conditions that led to the impasse and the result that resolved it — into a brand new production rule, added directly to procedural memory. Next time the same conditions arise, that new rule just fires, no subgoal, no search, the way a skill you’ve drilled enough stops requiring conscious deliberation. Newell treated this as SOAR’s entire theory of learning: all learning is chunking, arising as a side effect of doing (impasse-driven subgoaling), not a separate process bolted on.
This is a genuinely different design bet than lesson 10’s ACT-R. ACT-R keeps declarative memory — retrievable facts, each with a graded activation number that can be strong, weak, or anywhere between, decaying continuously over time (lesson 10’s B = −d·ln(t)) — as a distinct store from procedural memory (production rules, which either fire or don’t; no graded “activation” on a rule’s applicability the way there is on a fact’s retrievability). Classic SOAR has no such declarative store at all: everything, ultimately, becomes a production rule, all-or-nothing, compiled directly by chunking. Two architectures, two different bets about whether “remembering a fact” and “having a skill” are the same kind of thing underneath.
Given that contrast: which statement best captures the core difference?