In this paper, we investigate and experiment the notion of error correction memory applied to error correction in technical texts. The main purpose is to induce relatively generic correction patterns associated with more contextual correction recommendations, based on previously memorized and analyzed corrections. The notion of error correction memory is developed within the framework of the LELIE project and illustrated on the case of fuzzy lexical items, which is a major problem in technical texts.