Intelligent tutoring systems seek to optimize instruction and training by adapting and individualizing the learning experience on the basis of a student model (Shute, 1995). This model represents the system's estimate of the student's current knowledge or skill level, established from a performance history. Knowledge tracing (Aleven & Koedinger, 2002; Anderson, Conrad, & Corbett, 1989) is a dynamic, Bayesian approach to updating the estimates of probability of skill mastery in the student model. A fundamental shortcoming of this approach is that it does not include a representation of memory decay during periods of non-practice. As a result, traditional student modeling approaches are unable to make predictions regarding knowledge and skill changes under various future training schedules or to prescribe how much training will be required to achieve specific levels of readiness at a specific future time. In this paper, we propose a new knowledge tracing equation, computationally inspired by the learning and forgetting equations in the ACT-R cognitive architecture (Anderson et al., 2004), which uses performance history to baseline student model parameters and then extrapolates knowledge state transformation to predict future performance. We explore practical issues concerning predictive models of future trainee performance and the prescription of frequency and timing of optimal learning with training systems. For instance, we investigate how much data from the training history are necessary to achieve reasonable predictive validity, and we describe the impact of data granularity through a quantitative assessment of how adequately the model can fit and predict human performance curves across aggregate-level, team-level, and individual-level resolutions. The paper ends with a discussion of the implications of this research for the future of training and education.