Visual working memory (VWM) is a crucial part of our cognitive system. Currently there is an active debate how the apparent limitations of VWM should be described. Limited-slot and flexible-resource theories are discussed, but so far the temporal dynamics of representations stored in VWM are not fully understood. In this paper we present data that supports the notion of dynamic VWM contents with changing precision. To account for these observations in a qualitative way, we propose a neural network that is able to account for emerging capacity limits as well as for changes in the precision of stored information.