Two nonverbal short-term memory tasks, a Picture Recognition and a Paired Association Task, were performed by matched groups of aphasics and brain-damaged controls (N=52 each). For both tasks the alternatives on the answer forms differed systematically on three dimensions, one of which had been highlighted experimentally for each item before the target picture was indicated. In both tasks aphasics made more errors than braindamaged controls only with respect to those features that had been stressed beforehand; they were less able than controls to take advantage of focussed attention. The number of these errors was correlated with performance in the Token Test, a picture naming and a word comprehension task.