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.

--
Jeff Garzik
Bitcoin core developer and open source evangelist
BitPay, Inc.      https://bitpay.com/