The authors studied the operation of working memory in language comprehension by examining the reading of complex sentences. Reading time and comprehension accuracy in self-paced reading by college students were studied as a function of type of embedded clause (object-extracted vs. subject-extracted) and the types of noun phrases (NPs) in the stimulus sentences, including relative clauses and clefts. The poorer language comprehension performance typically observed for object-extracted compared with subject-extracted forms was found to depend strongly on the mixture of types of NPs (descriptions, indexical pronouns, and names) in a sentence. Having two NPs of the same type led to a larger performance difference than having two NPs of a different type. The findings support a conception of working memory in which similarity-based interference plays an important role in sentence complexity effects.