As far as I understand MAM users don't care if their transactions will be ever confirmed. They use the tangle just as a transport layer.
So my question is: Why would they use any tip selection algorithm at all and not just pick any two random transactions? Or, even worse, always select the same transactions creating disgusting clumps in the tangle?
To be honest, I am not sure whether you can be that careless about whether your transactions confirm.
There is no guarantee that unconfirmed transactions will (eventually) reach every node of the network. Also permanodes will probably not retain unconfirmed transactions forever (in case they reached the node originally). Also, it is hard to "timestamp" unconfirmed transactions (you cannot give an upper bound for the time, only a lower bound).
Therefore, if you want to use the tangle to store data that later should be proven to exist before (or to be proven that there are no other conflicting data), you would need your transactions to be confirmed, even if they are zero-valued. For other usescases, probably not.
