We develop a model capturing habit formation, projection bias, and present bias in an intertemporal-choice setting, and conduct a field experiment to identify its main parameters. We elicit subjects' pre- and post-treatment predictions of post-treatment gym attendance, using a habit-formation intervention based on Charness and Gneezy (2009) as an exogenous shock to treated subjects' gym preferences. Projection-biased subjects, projecting their current habit state onto their future expectations, will, ex-ante, under-estimate any habit-formation effect of our treatment. Naive present-biased subjects in both groups will overestimate their future attendance. Like Charness and Gneezy, we find subjects do form a significant short-run habit, though we find substantial decay caused by the semester break. Subjects appear not to embed this habit formation into their ex-ante predictions. Approximately one-third of subjects formed a habit equivalent to the effect of a $2.60 per-visit subsidy, while their predictions correspond to 90% projection bias over this habit formation. Moreover, subjects greatly over-predict future attendance, which we interpret as evidence of partial naivete with respect to present bias: they appear to expect their future selves to be two-thirds less "present biased" than they currently are.