In this talk, I’ll suggest a rethinking of the nature of visual working memory: first, I’ll argue that we have a separate object working memory and scene working memory system, which show themselves in a wide variety of tasks (ranging from developmental psychology to cognitive neuroscience) and each contribute to nearly all working memory tasks. Second, I’ll argue that working memory, like long-term memory, has a capacity that depends critically on the semantic meaning of what you are asked to remember, as opposed to being merely the persistence of perceptual representations. In particular, I’ll argue that one of the ways we recognize visual scenes is by treating the scene as a global texture and processing the distribution of orientations and spatial frequencies holistically across the entire scene, without recognizing any objects. I’ll show that we can see reflections of this texture-based scene processing pathway in visual working memory, suggesting that visual working memory consists of separate scene/texture and object representations. I’ll show evidence that even the simplest visual working memory experiment -- with just 3 colored dots -- actually relies on dissociable memory representations from the object system and scene system. In the second part of the talk, I’ll discuss some recent EEG work trying to understand the nature of visual working memory, and, in particular, how visual working memory is affected by knowledge. I’ll show that brain measures of how much is actively being stored in working memory demonstrate a greater capacity for real objects than for simple stimuli. This suggests that working memory has no fixed capacity -- instead, our ability to remember new information depends critically on our existing knowledge.