The problem with current tx replacement proposals is that it requires to trust a miner to replace non final tx in memory pool for final and commit it into a block, in other words it's just slightly more secured version of 0-confirmation transaction.

The idea is to have nLockTime transaction committed into a block, but it becomes spendable only after nLockTime expires and if at any point before that a final version of the transaction appears in a block then nLockTime transaction becomes invalid, needless to say between payer and payee, only payee should be able to finalize transaction.