Remember the Past: Distilling Datasets into Addressable Memories for Neural Networks