Do you deal with memory compaction yourself or let the models handle it?