We compared the accuracy of saccades made to seen or remembered visual targets in 3 patients with saccadic dysmetria due to cerebellar lesions. Saccadic dysmetria was worse for saccades to remembered targets and for saccades to flashed targets visible for only 150 msec (i.e., invisible at the time of saccade). Furthermore, no corrective saccades were made if the target was not visible. These results have implications for testing saccadic dysmetria and suggest new hypotheses about the control of saccadic amplitude.