Working memory (WM), the ability to momentarily maintain information in an active state, is central to higher cognitive functions. The processes involved in WM operate on a sub-second timescale, and thus evoked potential measures have an appropriate temporal resolution for studying them. In the experiment reported here, evoked potential covariances (EPC) between scalp recording sites were computed for a task requiring maintenance of numeric information in WM; these EPCs were compared to those observed in a control task which had the same stimuli and responses but less of a WM requirement. EPC patterns differed between conditions prior to the stimulus, and in an interval spanning the P300 peak in the match detection trials which required response inhibition. The pattern of prestimulus EPCs was more complex and left-sided in the WM task, when memory codes were being maintained and responses contingent on those codes were being prepared. P300 peak latency was 140 msec shorter in the WM task, and the P300 EPC pattern was more anterior and left-sided. In contrast, EPC patterns did not differ during early stages of stimulus processing or during response execution. These results suggest that distinct EPC patterns associated with WM only occur during intervals in which the information in an active state is being utilized for task performance.