You seem to be saying that it's not okay for me to lie even if I can't get caught, because then I'd be the "third-party beneficiary"

Right. "Third-party beneficiary" can be seen as a generalized action, where the action is to produce an agent, or cause a behavior of an existing agent, that works towards optimizing your value.

but it's somehow okay to lie and then erase my memory of lying. Is that right?

It's not okay, in the sense that if you introduce the concept of you-that-decided-to-lie, existing in the past but not in present, then you also have to morally color this ontological distinction, and the natural way to do that would be to label the lying option worse. The you-that-decided is the third-party "beneficiary" in that case, that distinguished the states of the world containing lying and not-lying.

But it probably doesn't make sense for you to have that concept in your ontology if the states of the world that contained you-lying can't be in principle (in the strong sense described in the previous comment) distinguished from the ones that don't. You can even introduce ontological models for this case that, say, mark past-you-lying as better than past-you-not-lying and lead to exactly the same decisions, but that would be a non-standard model ;-)