Many common activities rely on spatial knowledge acquired from nonvisual modalities. We investigated the nature of this knowledge by having people first look at a collection of objects on a desktop and then manually reconstruct the arrangement of objects without vision. Participants were required to reconstruct the display as though it had been rotated by 0°, 45°, 90°, 135°, or 180° relative to the view they could see. Performance on several measures of visual-spatial memory showed that participants had better visual memory for the view they had manually reconstructed than for the view they had studied visually for several minutes. These findings provide compelling new evidence that visual-spatial knowledge of very high fidelity can be acquired from nonvisual modalities, and reveal how visual and nonvisual spatial information may even be confused in the brain.