Visual landmarks introduce systematic distortions into spatial short-term memory for single target positions, the exact form of the distortion depending on the spatial layout of the landmarks. In two experiments, we investigated how the combined effect of two landmarks can be predicted from the effects of individual landmarks. Participants used a mouse cursor to reproduce the positions of briefly presented targets in the context of one, the other, or both landmarks. We found that distortions near a landmark are independent of whether another landmark is present, so that remembered space is partitioned into regions dominated by single landmarks. Interestingly, the display midpoint behaves like a "virtual landmark," with its own pattern of distortion. Results are inconsistent with current models of spatial memory distortions but suggest that attentional processes can lead to enhanced fidelity of salient regions in topographical neural networks while also introducing some spatial biases.