Research on working memory and language has followed two quite divergent paths. The first line of inquiry examines questions relating to the components and organization of working memory – whether there are specialized buffers, the nature of the link to long-term memory, and so on. For the most part, studies of this type have little to say about the workings of language per se – why it has the particular types of relative clauses or patterns of verbal agreement that it does, for example. These issues fall under the purview of a different line of research, which seeks to trace various fundamental properties of language to the cognitive processes involved in the storage and manipulation of information – working memory, broadly construed. The goal of the latter research program, which I will try to advance here, is to establish that general properties of working memory, however they are ultimately integrated into a theoretical model, can contribute to a deeper understanding of the human language faculty. I will focus here on three phenomena that help illustrate this point – a restriction on the interpretation of reflexive pronouns, a curious prohibition on phonological contraction in a type of wh question, and a baffling constraint commonly known as the ‘that-trace effect.’ A careful examination of their properties reveals a previously unsuspected finding: they are shaped by the need to minimize processing cost, a key factor in our understanding of working memory as well.