A configural theory of human amnesia is proposed. The theory predicts that amnesic patients will exhibit selective deficits on tasks that normal subjects perform by learning new configurations of stimulus elements. This prediction is supported by results for four amnesic patients who learned a nonconfigural control task but failed to learn the configural transverse patterning task even after extensive practice. Matched normal subjects easily learned both tasks. The theory provides unique and viable accounts of the central results in the human amnesia literature. Relations between the configural approach and other theories are discussed.