Translation memory software keeps every segment a team has had translated, with its translation per locale, and offers it back whenever a new release contains the same or a similar segment. What is translation memory in practice is a cost engine: the analysis of a release against the memory splits its word count into exact matches, fuzzy matches and new words, and the vendor charges a different rate for each, so a memory that holds well makes a release cost a fraction of its full rate. Translation memory systems come inside every translation management system and every translator's tool, and the question for a software team is not whether it has one but who owns it. This page sets out what the memory holds, how the leverage is worked, and why the translation memories should be the team's.
What the memory holds and how a match is scored
A translation memory holds source segments, usually a sentence or a string, paired with their translation in each locale, and the analysis of a new release compares each segment in it with the memory. A segment found verbatim, in the same context, is an exact match; one found with small differences, a changed number or one word, is a fuzzy match scored by similarity; one not found is new. The translation memory software does the scoring before any human sees the release, which is why the analysis report exists before the quote does.
Leverage is the share of the full cost the memory removed
The release's cost is the new words at the full rate, the fuzzy words at the fuzzy rate and the exact words at the review rate, per locale, and the leverage is how far below the full-rate cost that lands. The translation memory leverage sheet on this site works it from the analysis report's match shares and the vendor's rate card, with no account, and gives the cost per locale, the cost across every locale, the saving and the effective rate per word. It is the figure a team reads before accepting a quote, and the figure that decides whether a locale is affordable this release.
Why translation memories should belong to the team
A memory the vendor holds is the vendor's leverage: the team pays the rates the vendor sets against matches the team cannot see. A memory the team keeps, in its own translation memory system and exportable in TMX, is the team's: the analysis is read before the quote, a second vendor can be asked to quote against the same memory, and a change of vendor loses nothing. Xlifflane Pro keeps every translated segment against its key, so the memory grows with the catalog and stays with the team.
Questions people ask about translation memory software
What is translation memory, in one sentence?
The store of every segment a team has had translated with its translation per locale, used to match and price the segments in the next release.
Is translation memory the same as machine translation?
No. Machine translation generates a translation for a new segment; translation memory returns a translation a human already made for the same or a similar segment. Most tools use both, with the memory first.
What format does a translation memory export in?
TMX, the Translation Memory eXchange format, which any translation tool reads. A memory that cannot be exported in TMX is a memory the team does not own.