The most compelling neural evidence for working memory is persistent neuronal activity bridging past sensory cues and their contingent future motor acts. This observation, however, does not answer what is actually being remembered or coded for by this activity. To address this fundamental issue, we imaged the human brain during maintenance of spatial locations and varied whether the memory-guided saccade was selected before or after the delay. An oculomotor delayed matching-to-sample task (match) was used to measure maintained motor intention because the direction of the forthcoming saccade was known throughout the delay. We used a nonmatching-to-sample task (nonmatch) in which the saccade was unpredictable to measure maintained spatial attention. Oculomotor areas were more active during match delays, and posterior parietal cortex and inferior frontal cortex were more active during nonmatch delays. Additionally, the fidelity of the memory was predicted by the delay-period activity of the frontal eye fields; the magnitude of delay-period activity correlated with the accuracy of the memory-guided saccade. Experimentally controlling response selection allowed us to functionally separate nodes of a network of frontal and parietal areas that usually coactivate in studies of working memory. We propose that different nodes in this network maintain different representational codes, motor and spatial. Which code is being represented by sustained neural activity is biased by when in the transformation from perception to action the response can be selected.