In modern learning environments such as Ready Relevant Learning, often multiple learning systems work together (e.g. computer-based training, intelligent tutors, and training simulations). Effective recommendation has been well studied within each of these learning systems. However, when several systems are available that train in different ways, a new challenge emerges to understand the data that systems share, in light of their varied instructional designs and understanding of science of learning. Emerging data specifications and machine learning promise to help recommend which learning systems best fit individual learners’ needs and desired learning outcome. An exemplar recommendation component was created to drive training progression in several different learning systems. Recommendations were produced by combining and deconflicting learner information from multiple systems. Experiments with historical data and simulated students showed that the recommendation component could prioritize available learning systems and content adaptively. The recommender successfully inferred needed science of learning information, such as relating learning activities to skills and estimating the varying difficulty of skills. The initial research reported here focuses on objective learner performance metrics. Our results show that the recommender accurately matched ground truth in estimating learner mastery and skill difficulty. The recommender also incorporated simulated input from human instructors, which reduced its error rate to near zero. Finally, a simple exemplar algorithm deconflicted learner mastery estimates from different learning systems and used them to give learners qualitatively different recommendations. The feasibility demonstration reported here enables a 2018 human-participants study applying the same components to additional factors. Subjective learner states (boredom, confusion) and science of learning facts about training (welldefined vs. ill-defined, introductory vs. worked example, static vs. interactive) can drive the same recommendation tools (shared data specifications and machine learning). Our simulation studies suggest that recommendation across learning systems will make real training more effective than the sum of its parts.