Memory systems are on the verge of a renaissance: Scalable, persistent main memories (e.g., Intel’s 3DXPoint) are the first new technology to enter the upper layers of the memory hierarchy in 50 years. They bring a fundamentally new capability (i.e., persistence), a dramatic increase in capacity, and an array of complications (e.g., asymmetric read and write performance, power limitations, and wear out). This combination of characteristics raises a deceptively simple but fundamental question: What should we do with persistent main memory? In this talk, I will describe several potential answers and the systems my group has built to help understand how different answers affect performance, programmability, and other aspects of system design. I’ll also highlight the central challenges that these memories present and try to summarize what we have learned about them. Finally, I’ll describe what I see as the most interesting avenues for future work.