In applied settings, computational models of memory haveproven useful in making principled performance predictions.Specifically, historical data are used to derive modelparameters in order to enable out-of-sample predictions.Parameters are typically fit to meaningful subsets of data.However, labels that demarcate what constitutes a“meaningful” subset are not always available. Here, we utilizea data-driven method to cluster past performance into subsetspossessing statistical similarities. We contrast predictions fromcluster-specific model parameters with predictions based onsubsets that are artifacts of the experimental design. We showthat cluster-based predictions are at least as accurate as thechosen baselines and highlight additional advantages of thedata-driven approach.