Event-related potentials (ERPs) were employed to examine whether semantically opaque words were better remembered than were semantically transparent words because of the inconsistency between the meanings of an opaque word and its constituent characters. Experiment 1 employed lexical decisions as the encoding task. The opaque words were better remembered. Moreover, they elicited a greater left parietal ERP old/new effect, an index of recollection-based recognition, than did the transparent words. In Experiments 2 and 3, when semantic transparency judgment was employed as the encoding task to ensure that the constituent characters of the transparent words were processed elaborately to the same degree as those of the opaque words, the recognition advantage for the opaque words was reduced. The inconsistency between a whole word and its constituent characters rendered the opaque words more distinctive and thus better remembered than the transparent words.