Seems fairly "conceptually straightforward" (if not so easy): compress data (internal dependencies), use predictions (external dependencies), save states server-side, simply reduce garbage.