The question ‘what is working memory capacity?’ can be approached from the individual-differences and the experimental perspective. The first part of the article reviews individual-differences research that aims to determine the source of common variance among tests of working memory and related complex cognitive activities, such as reasoning and text comprehension. I propose that the ability to build, maintain and update temporary bindings between content and context representations underlies the shared variance of these variables. Multiple bindings must be maintained simultaneously for building new structural representations that underlie reasoning and language comprehension. Interference between these bindings limits the complexity of these structural representations. The second part of the article reviews experimental work, informed by computational modelling of the mechanisms of working memory. I introduce an interference-based model of working memory, SOB-CS, and review experiments with complex-span tasks that test predictions from that model. The experiments support the assumption that the capacity of working memory is limited by two forms of interference: interference by superposition of multiple bindings between distributed representations, and interference by confusion of similar items, and of items bound to similar contexts.