I'm curious why they conclude "we predict less change than we will actually experience" from this rather than "we recollect more change than we actually experienced."

The authors were aware of this issue and they addressed it twice in the paper. First, for study 1 (p. 96, column 3):

we compared the magnitudes of the predicted and reported personality changes in our sample to the magnitude of actual personality change observed in an independent sample of 3808 adults,

and found that the intraclass correlation of the reported personality changes in their sample matches the intraclass correlation of the actual personality changes in the independent sample. (This is consistent with memory being reliable, but it is not conclusive evidence.)

Second, (p. 98, column 1):

in study 3 we examined the end of history illusion in a domain in which memory was likely to be highly reliable. Rather than asking reporters to remember how extraverted they had been or how much they had once valued honesty, we asked them to remember simple facts about their strongest preferences, such as the name of their favorite musical band or the name of their best friend.

They found the same effect in study 3 as the one that they found in study 1 (suggesting that faulty memory is not the sole cause of the effect), but Ellenberg (which Qiaochu_Yuan linked to) has a critique of study 3 (as well as other parts of the paper).