The mechanism underlying our knowledge of spatial relationships was investigated. In the first experiment subjects were asked to examine an array of circles, located in different positions on a background; after removal of the array, they answered questions about the spatial relations between the circles. The array was constructed so that the number of circles and the number of relations between circles could be independently varied. The results showed that the number of correct answers decreased as the number of circles increased, but was unaffected by the number of relations between circles. This suggests that subjects learn the positions of the circles, from which relations between circles can be computed, rather than learning the relations directly. Evidence from the second experiment, using a similar experimental paradigm, suggests that the positions of the circles are learned by organizing them into higher-order units or groups.