I feel like a lot of this will be driven by implementation, and costs of changing the implementation. Additional look-backs are of course doable, but they incur some disk I/O costs. The fields available in memory for each mempool TX are
uint256 tx_hash; // hash of next field
CTransaction tx;
int64_t nFee; // Cached to avoid expensive parent-transaction lookups
size_t nTxSize; // ... and avoid recomputing tx size
int64_t nTime; // Local time when entering the mempool
double dPriority; // Priority when entering the mempool
unsigned int nHeight; // Chain height when entering the mempool
As a first pass, we may prune the mempool without additional db lookups quite easily based on time criteria. Or, additional in-memory indexes may be constructed to maintain hashes ordered by priority/fees.
Those techniques seem likely to be attempted before resorting to looking back two or three TXs deep at coin age -- which I admit is an interesting metric.