Distilling and Retrieving Generalizable Knowledge for Robot Manipulation via Language Corrections