Panic patients with agoraphobia were compared with normal controls on tasks of face recognition. The subjects were presented with 20 photos, and were required to make a judgement of the persons on the photos; shortly afterwards they were unexpectedly presented with a recognition task. In the first study, one task was to judge whether the persons on the photos were critical or accepting: unlike social phobics (Lundh and Ost, 1996b, Behaviour Research and Therapy, 34, 787-794), panic patients showed no bias for critical vs accepting faces on the recognition task. In a secondary study, the task was to judge whether the persons on the photos were 'safe' or 'unsafe', i.e. whether they could be relied on if the subject would need help in some situation. The results showed a recognition bias for safe vs neutral faces in panic patients. The index of recognition bias for safe faces correlated with avoidance of feared situations when accompanied by others, as measured by the Mobility Inventory. The possibility that memory bias in emotional disorders is a function of basic concern, or functional importance, rather than positive/negative valence is discussed. The results are also discussed in terms of degree of elaboration, exposure duration of the stimuli, and the generality of the findings.