The aim of this study was to determine the neuronal basis for memory impairment in Alzheimer's disease by taking advantage of the clinical and metabolic heterogeneity of this pathology. To this end, 19 patients satisfying the NINCDSADRDA criteria for probably Alzheimer's disease of mild-to-moderate severity underwent a detailed examination of the five memory systems according to Tulving's model, together with a PET measurement of resting regional cerebral glucose utilization (CMRGlc). Compared with controls, the patients as a group showed the expected memory and metabolic profiles of impairment. Correlations (corrected for the effects of ageing) were calculated between memory scores and CMRGlc (normalized by the vermis CMRGlc) using two methods: (i) the classic regions-of-interest method, based on a priori hypotheses and individual coregistered structural MRI; and (ii) the statistical parametric mapping method which allows a systematic voxel-by-voxel analysis, in a more descriptive and exploratory way. Significant correlations were above chance levels and largely consistent between the two methods. They were almost exclusively positive (i.e. in the neurobiologically expected direction) and their distribution showed striking differences according to each memory system. Thus, verbal episodic memory impairment was related to changes in a large neuronal network including not only the limbic structures (mesial temporal cortex, thalamus and cingulate gyrus, with left side predominance) but also the parietotemporal and frontal association cortices of the right hemisphere, possibly on a compensatory basis. Regardless of modality, short-term memory tests were mainly correlated with bilateral activity in posterior association cortex, and also with activity in left prefrontal cortex for the visuospatial span, possibly indicating essentially uniform strategies for the performance of the different tasks. As predicted, semantic memory scores correlated with activity in temporoparietal and frontal association cortices of the left hemisphere, and also with activity in left cingulate cortex. Thus, for episodic, short-term and semantic memory, many findings fit classical neuropsychology, while most of the less expected ones were consistent with recent results from functional neuro-imaging in healthy subjects, notably with the hemispheric encoding/ retrieval asymmetry (HERA) model; only few findings suggested possible reorganization processes and/or recourse to unexpected cognitive strategy. Finally, only negative correlations were found for perceptual priming and procedural memory; although they could arise by chance, some of these unexpected findings give rise to interesting hypotheses about the cognitive relationships between the most and least affected memory systems. This study documents the validity and usefulness of our approach in unravelling the neural substrates of cognitive impairment in brain pathology without focal tissue loss such as that seen in neurodegenerative diseases.