Word memorability is not unitary: the words that are easy to recognize are not always the words that are easy to recall. I asked what kind of lexical information each task draws on, using word-level recognition and free-recall memorability for 510 English nouns in public datasets. Three representations were compared under ridge regression with leave-one-out cross-validation: a 100-dimensional psycholinguistic norm space, a 300-dimensional distributed semantic space, and their combination. Distributed semantics contributed disproportionately to free recall, whereas the smaller recognition-specific contribution of the norm space was concentrated in lexical accessibility and familiarity. Six unsupervised indices of distributional neighborhood structure recovered most of the recall-specific advantage, whereas the same indices computed within the norm space did not. Human judgements showed a strikingly different structure. Participants' predictions of recognition and recall correlated .54 across words although the corresponding memory outcomes correlated only .07, and the task-specific representational asymmetry observed in memory was not recovered from their judgements. Thus, recognition and free-recall memorability draw on partly distinct lexical information, but this distinction is not reflected in people's beliefs about what will be memorable.