Different access modes to information stored in long-term memory can lead to different distributions of errors in classification tasks. We have designed a famous faces memory classification task that allows for the extraction of a measure of metric content, an index of the relevance of semantic cues for classification performance. High levels of metric content are indicative of a relatively preferred semantic access mode, while low levels, and similar correct performance, suggest a preferential episodic access mode. Compared with normal controls, the metric content index was increased in patients with Alzheimer's disease (AD), decreased in patients with herpes simplex encephalitis, and unvaried in patients with insult in the prefrontal cortex. Moreover, the metric content index was found to correlate with a measure of the severity of dementia in patients with AD, and to track the progression of the disease. These results underline the role of the medial-temporal lobes and of the temporal cortex, respectively, for the episodic and semantic routes to memory retrieval. Moreover, they confirm the reliability of information theoretic measures for characterizing the structure of the surviving memory representations in memory-impaired patient populations.