Amygdala function is of high interest for cognitive, social and psychiatric neuroscience, emphasizing the need for reliable assessments in humans. Previous work has indicated unsatisfactorily low within-subject reliability of amygdala activation fMRI measures. Based on basic science evidence for strong habituation of amygdala response to repeated stimuli, we investigated whether a quantification of habituation provides additional information beyond the usual estimate of the overall mean activity. We assessed the within-subject reliability of amygdala habituation measures during a facial emotion matching paradigm in 25 healthy subjects. We extracted the amygdala signal decrement across the course of the fMRI run for the two test-retest measurement sessions and compared reliability estimates with previous findings based on mean response amplitude. Retest-reliability of the session-wise amygdala habituation was significantly higher than the evoked amygdala mean amplitude (intraclass correlation coefficients (ICC)=0.53 vs. 0.16). To test the task-specificity of this finding, we compared the retest-reliability of amygdala habituation across two different tasks. Significant amygdala response decrement was also seen in a cognitive task (n-back working memory) that did not per se activate the amygdala, but was totally unreliable in that context (ICC~0.0), arguing for task-specificity. Together the results show that emotion-dependent amygdala habituation is a robust and considerably more reliable index than the mean amplitude, and provides a robust potential endpoint for within-subject designs including pharmaco-fMRI studies.