We propose that the earliest neuropsychological detection of Alzheimer's disease (AD) can be informed by current views about the neuropathogenesis of AD and cognitive models of memory and its neurobiological substrates. The primary impairment in early AD is encoding/consolidation, resulting from medial temporal lobe (MTL) pathology. On theoretical and empirical grounds, paired associate learning (PAL) appears to be the ideal paradigm for detecting MTL dysfunction in early AD. It has not been embraced as a test of choice, however, and this critical review discusses why the paradigm may have not fulfilled its potential. We suggest that a new PAL variant, 'associate-recognition', may prove to be clinically efficacious.