The separate banks will be for storing the assets (sample data) and the metadata respectively. So the actual audio files to play are kept in one bank. The metadata bank will contain all other information pertaining to the events and signal chains, including the output format.
The metadata banks are very small in comparison to the asset banks. If you don’t separate them out like this, then both the assets and metadata are bundled into the same bank, meaning for each platform you will have the same assets present in each platform’s bank which will take up a lot of space. So you can have one asset bank that that is shared across multiple platform metadata banks.
Taking the Examples project as an example - having metadata and assets in the same bank will result in each platform being 27.3MB, a total of 81.9MB for 3 different mix banks. However, with separated assets and metadata banks, this would be 115KB for each group of platform metadata banks, plus 27.2MB for the shared asset banks - a total of 27.3MB. So for the same space of one platform’s group of bundled banks you can have three mixes.
In game (regardless of Unity or Unreal) you would only need to load the asset banks once needed but you can load and unload the different mixes when required. If an event instance is already in play, it will be destroyed once the bank is unloaded, so it would be advisable to only allow changing mixes from a settings menu before starting the game. You could have a system in place that memorizes all playing event instances and their timeline positions & parameter values and restart all of them if the mix bank is changed.
I hope this helps.