The transverse patterning (TP) task (A+ B−, B+ C−, C+ A−) has played a central role in testing the hypothesis that medial-temporal (and, in particular, hippocampal) brain damage selectively impairs learning on at least some classes of configural (i.e., nonlinear) learning tasks. Results in the animal and human literature generally support that hypothesis. Reed and Squire [Impaired transverse patterning in human amnesia is a special case of impaired memory for two-choice discrimination tasks. Behavioral Neuroscience, 113, 3–9, 1999], however, advanced an alternative account in which impaired TP performance in amnesia reflects a generic scaling artifact arising from the greater difficulty of the TP task compared to the elemental (i.e., linear) control task that is typically used. We begin with a critique of Reed and Squire, countering their conceptual arguments and showing that their results, when analyzed appropriately, support the configural deficit hypothesis. We then report results from eight new amnesic patients and controls on an improved version of the TP task. Despite substantial practice, accuracy of patients with bilateral hippocampal damage due to anoxia reached and maintained an asymptote of only 54% correct, well below the maximum accuracy obtainable (67%) in the absence of configural learning. A patient with selective bilateral damage to the anterior thalamic nuclei exhibited a TP accuracy asymptote that was near 67%, a pattern of two out of three correct consecutive trials, and a pattern of nearly always answering correctly for two of the three TP item pairs. These results are consistent with a set of unique and parameter-free predictions of the configural deficit hypothesis.