The accelerating effects of adaptive training systems are well established (Lesgold, 2012; Cohn & Fletcher, 2010). This power might be enhanced further by scheduling training to accelerate acquisition and scheduling re-training to reduce decay. Models of acquisition and decay have been available to support scheduling since the early days of memory research (Ebbinghaus, 1913). But these models are derived mainly from laboratory tasks that are learned and executed over seconds or minutes, and performed in isolation from competing tasks. The models are much less explanatory or predictive over real world tasks that are complex, learned and executed over hours and days, and situated in a river of daily assignments that impose the scientifically acknowledged cause of skill decay: interference with memory retrieval (Farr, 1987; Arthur, et al., 1998). In this paper, we propose a new approach to acquisition and decay modeling to make the science of skill acquisition and decay more useful and usable. The approach applies machine learning techniques to model skill acquisition and decay. We apply these methods to a large dataset from a game-ified working memory exercise, compare the performance of these methods with a conventional technique, and present the argument for applying these methods to predict learning and schedule training for realistically complex tasks such as system diagnosis and corrective maintenance.