Rewind: Some points from our research

There are lots of projects out there that are playing with voxels. Only a few have achieved demonstration state, and most of those seem flawed in some important way. Minecraft seems to be the cleanest implementation so far. I tried running Manic Digger, but it suffered performance problems, kept eating memory to fill 2GB, then hung - looking at the source code, it doesn't use SVOs (the special Octtree structure I worked out before the first Vol post), but instead caches uncompressed chunks of the landscape.

From my basic research, I learned that one needs:

1. OctTrees (or similar) to store a 3D landscape of voxels.
2. To discard parts from memory when they're no longer needed.
3. A procedural landscape generator, which is repeatable, and doesn't depend on what's going on elsewhere - this is limiting, but I can see ways around it.
4. [inferred] Given that the procedural aspect can be generated at any time, we have freedom to discard it as required, because we can always regenerate it on demand. We want a player to be able to dig and build in the landscape, so we'll need to store modifications in a separate octtree. The actual state of the landscape can be ascertained by reading the procedural content, and then overlaying the modification content. Modifications can be paged out as necessary, but never deleted.
5. Using conventional (straightforward) programming, adding attributes to voxels will drastically increase the memory usage for ALL voxels. We need to find a way around this (and we think we have).
6. People seem to think that a multiplayer is much more fun, and most developers have prioritied a networked client-server topology, though I've only seen 'shared enviroments' rather than any true player-to-player interaction or co-op missions. I think I'll ignore this for now, but try not to close the door to this option.