Working memory (WM), the ability to briefly retain and manipulate information in mind, is central to intelligent behavior. Here we take advantage of the high temporal resolution of electrophysiological measures to obtain a millisecond timescale view of the activity induced in distributed cortical networks by tasks that impose significant WM demands. We examined how these networks are affected by the type and amount of information to be remembered, and by the amount of task practice. Evoked potentials (EPs) were obtained from eight subjects performing spatial and verbal versions of a visual n-back WM task (n = 1, 2, 3) on each of three testing days. In well-trained subjects, WM tasks elicited transient responses reflecting different subcomponents of task processing, including transient (lasting 0.02-0.3 s) task-sensitive and load-sensitive EPs, as well as sustained responses (lasting 1-1.5 s), including the prestimulus Contingent Negative Variation (CNV), and post-stimulus frontal and parietal Slow Waves. The transient responses, with the exception of the P300, differed between the verbal and spatial task versions, and between trials with different response requirements. The P300 and the Slow Waves were not affected by task version but were affected by increased WM load. These results suggest that WM emerges from the formation of a dynamic cortical network linking task-specific processes with non-specific, capacity-limited, higher-order attentional processes. Practice effects on the EPs suggested that practice led to the development of a more effective cognitive strategy for dealing with lower-order aspects of task processing, but did not diminish demands made on higher order processes. Thus a simple WM task is shown to be composed of numerous elementary subsecond neural processes whose characteristics vary with type and amount of information being remembered, and amount of practice.