Recall from visual working memory shows averaging affects. For example, the recalled position of a memorised item is biased toward the average location of all items in a memory array. A recent suggestion is that averaging reflects an attempt to optimise single-item recall by exploiting ensemble statistics. This proposal predicts that the average location is memorised independently from that of individual items. We compared normal subjects’ perceptual estimates of the centre of mass (COM) of three-stimulus dot arrays, COM from recall and single items from recall. Perceptual estimates of COM showed a systematic bias toward the array’s incenter, COM recall did not show this bias. The precision of COM recall was lower than COM perceptual estimates and higher than single item recall. In a right hemisphere patient with left hemianopia and neglect, COM perceptual estimates were systematically biased contralesionally, while COM recalls were biased ipsilesionally, confirming the dissociation between perception and recall. These findings suggest that COM is recalled by averaging the memorised items’ positions rather than by retrieving its memorised perceptual estimate. Averaging in spatial recall may arise instead from a reference frame transformation, ensuring that the relative position of the item in the sample array is recalled.