Two experiments were conducted to study selective memory bias favoring anxiety-relevant materials in patients with anxiety disorders. In the 1st experiment, 32 patients with generalized anxiety disorder (GAD), 30 with social phobia (speaking anxiety), and 31 control participants incidentally learned GAD-relevant words, speech anxiety-relevant words, strongly pleasant words, and words with a neutral valence. Participants did not show any explicit memory bias for threatening materials. Thirty patients suffering from panic disorder (PD) with agoraphobia and 30 controls took part in the 2nd experiment. The design was similar to the 1st experiment. This time a highly specific selective memory bias for threatening words was found. Words describing symptoms of anxiety were better recalled by PD patients. Results are consistent with previous findings but are inexplicable by existing theories.