Background: Students increasingly use generative AI in medical education, yet most interactions occur through generic chatbots with limited educational scaffolding. We developed a grounded AI learning environment designed to support self-directed learning using evidence-informed educational principles. Methods: First-year medical students participating in a two-week extracurricular summer learning module were randomized to traditional independent study or access to a grounded AI learning agent. The agent was grounded in curated hematology content aligned with future medical training and incorporated metacognition, adult learning, coaching, and assessment design to scaffold planning, active learning, and formative self-assessment. Participants completed study trackers, post-study surveys, and a curated, content-aligned multiple-choice assessment from a commercial medical education and licensure preparation platform. Analyses were descriptive and hypothesis-generating. Results: AI access alone did not consistently explain assessment performance. Students varied substantially in how they integrated the grounded AI into their learning. Higher-performing learners more frequently documented retrieval practice, reasoning, feedback seeking, adaptive learning, and multimodal study behaviors. Survey responses demonstrated that learners used the AI as a tutor, coach, planner, and feedback partner rather than simply an information source. Conclusions: Educationally grounded AI may provide greater value than generic conversational AI by supporting meaningful learning behaviors. This pilot highlights the importance of educational architecture and learner engagement as key considerations for future AI-assisted medical education.