The problem is that they are not temp, but permanent files. Steinberg has acknowledged the issue and promised a solution in 11.0.20, but that didn’t materialise.
The thread I refer to goes into some detail (Cubase 11 ARA folder is gigantic from using Melodyne) with lots of examples, but the short version is:
If you loop record for instance vocals with say 5 takes. You then comp that vocal using parts from one or all of the takes. Each time you then select a part of a take to edit that in Melodyne, Cubase doesn’t just make a copy of that part of the take, it doesn’t even just make a copy of that take, but it makes a copy of all 5 takes. That’s fine if it’d stop at that. But the next section you edit in Melodyne, can be from the same take as the previous take or any of the other 5 takes it makes another copy of all 5 takes.
Why doesn’t Cubase just make a copy of the part of the take that you’re currently editing, or at least just a copy of the take of which you’re editing a small piece? That would be logical to me. At least, if Cubase needs to make a copy of all 5 takes, at least reuse that copy for the other edits I want to make. There is no reason Cubase need to make another copy of all five takes just to edit 1 bar in a 16-bar loop. You end up with 79 bars of audio that you don’t need, for each 1 bar edit you make. Do an edit in each bar and you’ll end up with 1264 bars ((16*5-1)*16) of unused audio.
This may not sound like a problem, but when you have a typical 3 min pop song with a lead a few doubles and a number of harmonies, totalling say 20 tracks of vocals. All of a sudden you have an ARA folder that is 30-50GB, but the underlying audio folder may just be 1GB. Not only does that ARA folder becomes gigantic, but Cubase also isn’t able to handle it and even the most powerful computers grind to a halt, to the extent that something so simple as selecting an event takes many seconds or stop playing the project takes several minutes.
As you can see, it is unusable. And as a result, all Melodyne edits need to be printed and the original edits deleted/archived to be able to work with it. It makes workflows like these (Comping with Melodyne (ARA) - YouTube) impossible for anything but demonstration in a controlled environment.
You may say that you should commit your edits. True, but when you are in the middle of the creative process you shouldn’t have to think about that when you’re comping vocals, as you very quickly run into trouble if you’re not careful and print every single edit you do.
This makes the ARA implementation somewhat redundant, as it nothing more than an inline destructive workflow.