rEDMRec: Distilling Large Language Model Reasoning into an Editable Experience Memory for Recommendation