The concept of a translation memory has been around for a long time--more than twenty years--but only recently has it become a significant commercial entity. Basically, a translation memory is a system which scans a source text and tries to match strings (a sentence or part thereof) against a database of paired source and target language strings with the aim of reusing previously translated materials. Some translation memories attempt only literal matching, ie can only retrieve the exact match of a sentence, while others employ fuzzy matching algorithms to retrieve similar target language strings, flagging differences. The flexibility and robustness of the matching algorithm largely determine the performance of the system, although for some applications (ie, highly repetitive material) the recall rate of exact matches can be high enough to justify the literal approach.
Translation memories are typically integrated into translation workstation packages, where they can be used in tandem with a terminology management system, a multilingual dictionary, and even raw MT output.