Research on memory may benefit from paradigms that permit graded characterization of memory performance, but a simple variance-based approach to the analysis of such graded data confounds two potential sources of error: the probability of memory and the fidelity of memory. Such data are more properly modeled by a mixture distribution, thereby permitting explicit estimation of both the probability and fidelity of memory. An expectation-maximization algorithm is presented for fitting such data to a mixture model, and Monte Carlo validation of this tool reveals circumstances under which it may be expected to be most effective. Limitations of the tool are outlined with respect to potential confounds in experiment design and interpretation of results. Finally, approaches to ameliorating such confounds are discussed. An R procedure for fitting response error data to mixture models may be downloaded from http://brm.psychonomic-journals.org/content/supplemental.