I propose that the capacity of working memory places a specific limit on the maintenance of temporary bindings. Two experiments support this binding hypothesis: Participants remembered word lists of varying length. When tested on a randomly selected word, their error rates increased with the length of the list, reflecting a limited capacity for short-term maintenance. This increase in errors was predominantly due to binding errors: People confused the correct word with other words of the current memory list, but very rarely with words not in the list. The frequencies of response choices were analyzed through two measurement models - one based on the assumption of discrete memory states, one on the assumption of continuous memory strength - that capture memory for items and for bindings in separate parameters. Increasing memory set size impaired binding memory but not item memory, supporting the binding hypothesis.