Dual-process models differentiate between two bases of memory, recollection and familiarity. It is routinely claimed that deeper, semantic encoding enhances recollection relative to shallow, non-semantic encoding, and that recollection is largely a product of semantic, elaborative rehearsal. The present experiments show that this is not always the case. In four experiments, the rhyme recognition test was adapted to two popular assessments of recollection (the Remember-Know technique and the process-dissociation procedure). The rhyme recognition test provides a better match to a non-semantic (phonological) encoding condition than to the semantic encoding condition. The experiments revealed a consistent reversal of the usual levels-of-processing effect, such that the measures of recollection were higher for the non-semantic than semantic encoding condition (the familiarity measures registered no differences between encoding conditions). This indicates that unqualified statements about particular encoding conditions producing recollection are not well founded. More generally, the results underscore the cue-dependent nature of recollection and transfer-appropriate-processing analyses of recollection. (Contains 5 tables.)