One of the big items which reduced memory usage was moving away from representing the interpreter instructions in a tree and instead storing them as an actual bytecode. Now, each instruction follows each other in memory and does not involve pointer-chasing.