This article proposes a hierarchical learning architecture for safe\ndata-driven control in unknown environments. We consider a constrained\nnonlinear dynamical system and assume the availability of state-input\ntrajectories solving control tasks in different environments. In addition to\ntask-invariant system state and input constraints, a parameterized environment\nmodel generates task-specific state constraints, which are satisfied by the\nstored trajectories. Our goal is to use these trajectories to find a safe and\nhigh-performing policy for a new task in a new, unknown environment. We propose\nusing the stored data to learn generalizable control strategies. At each time\nstep, based on a local forecast of the new task environment, the learned\nstrategy consists of a target region in the state space and input constraints\nto guide the system evolution to the target region. These target regions are\nused as terminal sets by a low-level model predictive controller. We show how\nto i) design the target sets from past data and then ii) incorporate them into\na model predictive control scheme with shifting horizon that ensures safety of\nthe closed-loop system when performing the new task. We prove the feasibility\nof the resulting control policy, and apply the proposed method to robotic path\nplanning, racing, and computer game applications.\n