One aspect of metacognition is the monitoring of memory or comprehension measured with retrospective confidence judgments after test taking. The research questions of the present study were whether different measures for the accuracy of such confidence judgments are stable over learning time, whether they generalize over two different tests, and whether they predict the learning outcome. In order to answer these questions, a study was conducted in which university students (N = 113) learned about the basic concepts of operant conditioning for 30 min. Knowledge tests with confidence judgments presented after each item were obtained before learning, after 10 min, and at the end of the learning session. Bias and absolute bias were calculated as absolute measures of accuracy, and gamma, Pearson's r, and d(a) were calculated as relative measures of accuracy. The results showed that the absolute measures were stable, but that the relative measures were not. Furthermore, absolute bias, gamma, and Pearson's r obtained 10 min after the beginning of the learning process predicted the learning outcome. The results are discussed with regard to research on different measures of accuracy for confidence judgments.