In this talk, I discuss my work on load balancing in MADNESS. Computational work in MADNESS takes the form of an octree with multiwavelet coefficients at each node, so many traditional load balancing techniques, designed for finite element domain decomposition or particle simulations, do not work well. We instead propose the melding algorithm, a recursive process in which leaf children are assimilated into their parent and then a depth-first partitioning of the tree is performed, resulting in an array of choices for distributing the octree across the machine. I discuss the motivation for developing the melding algorithm, describe the method and how it works, including ways to determine which configuration is best based on the target machine, and discuss its theoretical advantages over alternative load balancing techniques. Finally, I present preliminary results of using the melding algorithm on ORNL's Jaguar Cray XT4 machine.