In two experiments, we investigated the comprehension of sentences with repaired NPs. In our first experiment, we applied an acceptability task after speeded auditory presentation of French versions of sentences with and without repairs like I will go to (the butcher, uh no,) the baker. I need some bread/meat. While repairs led to reduced acceptability for consistent continuations, the inconsistent continuation was more acceptable when a compatible but repaired constituent had been mentioned before, suggesting that the to-be-repaired constituent was not fully overwritten by the correction. In our second experiment, we auditorily presented participants with the stimuli compiled for Experiment 1, while they looked at corresponding visual stimuli. Evidence from eye fixation patterns suggests that the to-be-repaired constituent was actually suppressed online during sentence processing. We suggest that the acceptability judgments are mainly the result of offline reconstruction of memory traces following Gimenes et al., (2009).