This seems to be an optimization problem. Certain statistics will probably predict the probability of a UTXO being used in the next x days again like amount, age, prevout, etc and maybe those UTXOs above a threshold probability could have a favorable spot in memory and others live on disk. This would have to be a very heuristic and rule-based approach, statistical models would have to be kept up to date etc... which iss all not really ideal. Maybe there's a better way?

The real fix seems to be utreexo which feels a few years away still.