Reasoning reuse: caching how the model thinks, not just what it says
Chain-of-thought tokens are the most expensive ones you buy. Crowkis extracts the thought's skeleton, abstracts the specifics, and recomposes it for the next input that shares its shape.
Response-level caching has a ceiling: it only saves when the final answer transfers. But look at where the tokens actually go in hard queries — the reasoning. Step-by-step derivations, plan decompositions, structured analyses. Different inputs constantly share identical reasoning shapes with different specifics, and response caching can't see the kinship.
Crowkis's reasoning store parses chain-of-thought output into typed steps, abstracts the particulars — numbers, dates, entities — into slots, and signs the step-type sequence. That signature is the reasoning's fingerprint: when a new query's shape matches a stored skeleton, the recomposer substitutes the new slots into the proven structure instead of re-purchasing the derivation.
Reuse only when meaning, structure, confidence, and trust all agree.
The worked-example case makes it concrete: the solution structure for one amortization calculation is the solution structure for all of them. First solve costs full reasoning tokens; every structural sibling after costs a recomposition. Math help, troubleshooting trees, policy analyses — anywhere thinking has a repeatable shape, the shape is now an asset.
The bottom line
It ships in every edition, gated by the same confidence machinery as everything else — a skeleton only serves where the match clears the bar. The deepest savings in the product, hiding in the step between question and answer.