As expected, this creates too much memory fragmentation. Trouble is, we will have to re-implement all BaseN-encoding on top of an alternative representation (see also #7), as otherwise we'll end up incurring a memcpy in almost every place multihashes are used.
