This study demonstrated that semantic transparency as a linguistic property modulates the recognition memory for two-character Chinese words, with opaque words (i.e., words whose meanings cannot be derived from constituent characters-e.g., "[/guang/, light][/gun/, stick]", bachelor) remembered better than transparent words (i.e., words whose meanings can be derived from constituent characters-e.g., "[/cha/, tea][/bei/, cup]", teacup). In Experiment 1, the participants made lexical decisions on transparent words, opaque words, and nonwords in the study and then engaged in an old/new recognition test. Experiment 2 employed a concreteness judgment as the encoding task to ensure equivalent semantic processing for opaque and transparent words. In Experiment 3, the neighborhood size of the two-character words was manipulated together with their semantic transparency. In all three experiments, opaque words were found to be better remembered than transparent words. We concluded that the conceptual incongruence between the meanings of a whole word and its constituent characters made opaque words more distinctive and, hence, better remembered than transparent words.