The trick is that there is an algorithm that describes the world so nothing needs to be stored until it is modified, and even then those changes are limited to

The trick is that there is an algorithm that describes the world so nothing needs to be stored until it is modified, and even then those changes are limited to a certain view distance. On the cpu side, a much lower resolution voxel data set is used for collision, pathfinding, etc (Also local to the camera)