Start with the part almost no one says out loud: a genuine expert usually cannot fully explain what they just did. Ask a fluent driver to narrate every decision in a highway merge, or a skilled diagnostician to list every signal that led to a correct call, and the honest answer trails off into "I just knew." This is not a communication failure. It is the documented, expected endpoint of learning a skill well -- named directly by the philosopher and scientist Michael Polanyi in 1966: "we can know more than we can tell."^[1]

Cognitive science gave that endpoint a mechanism, not just a phrase, the following year. Paul Fitts and Michael Posner's 1967 three-stage model of skill acquisition traces exactly how a skill gets there: the cognitive stage, where a learner consciously works through explicit steps one at a time; the associative stage, where practice smooths those steps into fewer errors and more consistency; and the autonomous stage, where the skill becomes proceduralized and largely automatic -- fast, reliable, and no longer dependent on the conscious, step-by-step access that built it in the first place.^[2] The steps were real and necessary. They were also always meant to be temporary.

A separate, independent line of memory research explains why the steps specifically are what disappear. Fuzzy-trace theory, developed by Charles Brainerd and Valerie Reyna, holds that a mind keeps two independent kinds of memory trace for the same experience: a verbatim trace, recording exact surface detail, and a gist trace, recording the underlying meaning or shape.^[3] Verbatim traces decay fastest and are the most vulnerable to interference from other similar experiences -- practice the same category of problem many times, in slightly different forms, and the specific steps of any one instance become harder to isolate, not easier. Gist traces do the opposite: they persist, and unlike verbatim detail, they are what actually transfers to a new problem that doesn't look exactly like the ones already practiced.^[3]

This is not a reason to skip the steps -- it is a reason not to mistake reciting them for the goal. Research directly comparing procedural and conceptual instruction bears this out with real students, not just theory: children taught a memorized procedure could reliably solve the exact type of problem they were drilled on, but instruction built around the underlying principle produced significantly more transfer to problems the students had never seen -- even though both groups learned to execute a correct procedure at roughly the same rate.^[4] Procedural and conceptual knowledge actually develop together, each reinforcing the other, rather than one simply replacing the need for the other.^[4] The steps are the substrate. The shape built from repeating them is the part that generalizes.

This argument also has a real limit, and stating it precisely matters more than smoothing it over. Everything above describes what happens to knowledge that starts as something learnable secondhand -- a procedure, a description, an instruction -- and gets refined through repetition into a transferable shape. Not all knowledge starts that way. Bertrand Russell drew a sharper line in 1910: knowledge by acquaintance, direct and unmediated awareness of a thing itself, versus knowledge by description, built from characterizations of a thing rather than contact with it -- and argued that descriptive knowledge, however refined, ultimately depends on a foundation of direct acquaintance it cannot manufacture on its own.^[6] No number of repeated encounters with a description of the ocean produces the acquaintance of having smelled salt air. Gist-formation makes secondhand knowledge more durable and more transferable than the verbatim version of the same secondhand knowledge -- it does not convert secondhand knowledge into firsthand knowledge. Overconfidence about what a well-formed shape actually contains, mistaking a durable description for direct contact with the thing it describes, is its own separate failure this piece is not a defense against.

The actual implication is narrower and more useful than "teach concepts instead of procedures." It is: build instruction and evaluation around the stage a skill is actually supposed to reach, not the stage that is easiest to test. A fill-in-the-blank test of memorized steps measures the cognitive stage -- the one every framework here agrees is temporary, the fastest-decaying, and the most vulnerable to being overwritten by the next similar problem. What a mind actually keeps, uses on problems it has never seen, and eventually can no longer even explain, is the shape underneath. That is not a failure to retain the lesson. By every framework named here, it is what successfully learning the lesson was always supposed to look like.

A related case, on the other side of the same gap: this site has also traced a real, dated case where an engineer had a correct, fully explicit, fully articulated warning -- and it still couldn't move a room already committed to a schedule. Explaining well and being believed are not the same accomplishment.