We tested whether visual stimulus material that is assumed to be processed in different cortical networks during perception (i.e., faces and spatial positions) is also topographically dissociable during long-term memory recall. With an extensive overlearning procedure, 12 participants learned paired associates of words and faces and words and spatial positions. Each word was combined with either one or two positions or one or two faces. fMRI was recorded several days later during a cued recall test, in which two words were presented and the participants had to decide whether these were linked to each other via a common mediator, i.e., a face or a position. This paradigm enforces retrieval from long-term memory without confounding recall with perceptual processes. A network of cortical areas was found to be differently activated during recall of positions and faces, including regions along the dorsal and ventral visual pathways, such as the parietal and precentral cortex for positions and the left prefrontal, temporal (including fusiform gyrus) and posterior cingulate cortex for faces. In a subset of these areas, the BOLD response was found to increase monotonically with the number of the to-be-re-activated associations. These results show that material-specific cortical networks are systematically activated during long-term memory retrieval that overlap with areas also activated by positions and faces during perceptual and working memory tasks.