Sampling-based planners are effective in many real-world applications such as robotics manipulation, navigation, and even protein modeling. However, in the absence of any prior information, sampling-based planners are forced to explore uniformly or heuristically, which can lead to degraded performance. One way to improve performance is to use prior knowledge of environments to adapt the sampling strategy. In this thesis, we propose Pyre, a retrieval-based framework for learning sampling distributions for high-dimensional motion planning. Pyre decomposes the workspace into local primitives, memorizing local experiences by these primitives in the form of local samplers, and storing them in a database. Pyre synthesizes a global sampler by learning to retrieve local experiences relevant to the given situation. This method transfers knowledge between diverse environments that share local primitives and significantly speeds up planning. Pyre learns incrementally and significantly improves performance with only a handful of examples, achieving better generalization over diverse tasks and environments as compared to prior approaches. We demonstrate the effectiveness of Pyre on 2D and 3D environments as well as with a simulated and real Fetch Robot tasked with challenging pick-and-place manipulation problems. Additionally, in this thesis, we also contribute MotionBenchMaker, an open-source benchmarking tool for realistic robot manipulation problems. It is an extensible, easy-to-use tool that allows users to both generate datasets and benchmark them by comparing motion planning algorithms.