Williams, Watts, MacLeod and Mathews' (1988) [Cognitive psychology and the emotional disorders. Chichester, Wiley] model of anxiety and cognition leads to the prediction that anxious subjects will show an implicit, but not an explicit, memory advantage for threat-related information. Mathews, Mogg, May and Eysenck (1989) [Journal of Abnormal Psychology, 98, 401-407] obtained marginally significant support for this prediction in an experiment that tested memory using word stem completion tasks following a self-referent encoding procedure. However, neither the reliability nor generality of these findings have been established. The current experiment was designed to provide a conceptual replication of Mathews et al.'s study, using different tests of implicit memory (i.e. tachistoscopic identification) and explicit memory (i.e. recognition) and an alternative type of encoding task (i.e. colour naming stimulus words). 16 generalised anxiety disorder patients, and 16 non-anxious control subjects were tested. As predicted, the anxiety patients showed a relative implicit memory advantage for threat-related stimulus words, while the two subject groups did not differ in their pattern of explicit memory performance. These results support the predictions generated by Williams et al.'s model of anxiety and cognition.