In a previous investigation, we raised the hypothesis that in Alzheimer's disease the cerebral structures implicated in episodic memory deficits may differ according to the severity of cognitive impairment. To test this hypothesis, Story Recall test scores and PET measurements of resting cerebral glucose utilization, a measure of synaptic integrity, were obtained in 40 patients. Using SPM96 (statistical parametric mapping 1996), positive correlations between the two sets of data were calculated on a voxel basis, first in the whole patient sample and then separately in the two subgroups of 20 patients differing in Mini-Mental State Examination score, i.e. those with least impaired and those with most impaired performance ('less severe' and 'more severe' subgroups, respectively). In the whole sample, significant correlations (P < 0.05, corrected for multiple tests) involved bilaterally not only several limbic structures (the hippocampal/rhinal cortex regions, posterior cingulate gyrus and retrosplenial cortex) but also, and less expectedly, some temporo-occipital association areas. However, the subgroup analysis disclosed that, in the less severe subgroup, all significant correlations (P < 0.005, uncorrected) were restricted to the parahippocampal gyrus and retrosplenial cortex, in accordance with both the distribution of changes in tau in early Alzheimer's disease and the known involvement of this network in normal and impaired memory function, while in the more severe subgroup they mainly involved the left temporal neocortex, which is known to be implicated in semantic memory. These findings suggest that, when episodic memory is mildly impaired, limbic functions are still sufficient to subserve the remaining performance, whereas with more severe memory deficit resulting from accumulated pathology the neocortical areas that are normally involved in semantic memory are recruited, perhaps as a form of (inadequate) compensatory mechanism.