What a translation memory is and how to start using one

What a translation memory is and how to start using one: a translation memory keeps every sentence you translate next to its translation, and offers it back when the same or a similar sentence comes again. Here is how it works, what it doesn’t do, and how to start one from the work you already have.

What a translation memory is

A translation memory, or TM, is a database of pairs: a sentence in the source language and the translation you confirmed for it. Each pair is called a translation unit. CAT tools such as Trados, memoQ, OmegaT and Lexora split a document into segments, usually sentences, and look each one up in the memory before you translate it.

A memory belongs to a language pair, such as English to German. Most translators keep one for each client, so one client’s terms and style never turn up in another’s work.

Memories travel between tools as TMX files, an XML format made for exchanging translation memories. If a client sends you their memory, it will usually be a TMX file.

How matches work

Exact match. The same sentence was translated before. You read it and confirm it.
Fuzzy match. A similar sentence was translated before. The tool shows how similar as a percentage and marks what differs, and you change only that.
Repetition. The same sentence appears more than once in the project. You translate it once.
Concordance. A search of the memory for a word or phrase, to see how you translated it before.

Each tool works the percentage out its own way, so a 75% match in one tool is not the same as in another. In Lexora it is the share of characters that stay the same between the two sentences, counted without differences in capitals, apostrophes and spacing. Two sentences that differ only in those score 99%, never 100%, so a real difference never passes as an exact match.

Many agencies pay less for matched words, by bands of match percentage. That is why their analysis counts words by their best match, and why your own memory affects what you can quote.

What a translation memory doesn’t do

It doesn’t translate. It gives back what a person already translated. Machine translation and AI write something new; a memory only remembers.
It doesn’t know if you were wrong. A slip you confirm comes back every time the sentence does. In Lexora, undoing a confirmation removes its memory entry too, so a slip you catch at once doesn’t stay.
It doesn’t understand context. The same English sentence can need two translations in two documents. Read an exact match before you confirm it.

How to start using one

Make one memory for each client, and name it after the client, so new projects for them find it.
Turn old work into memory. If you have earlier documents and their translations but no memory, align them: the tool pairs the sentences of the two files, and you check the pairs it is unsure of before they go in.
Ask your clients for theirs. Agencies and direct clients often have a memory for their texts. Ask for it as TMX.
Bring what you have from other tools. Export from your old tool as TMX, or, in Lexora, import Trados .sdltm memories directly.
Confirm as you go. Every sentence you confirm goes into the memory and comes back for the next job.
Keep a copy. Export your memories as TMX now and then and keep the files somewhere safe. A memory built over years is worth more than any one job.

In Lexora, the help article on memories and glossaries shows each of these steps, and the translation memory software page lists what Lexora reads and where it keeps your memory.

Who owns the memory

It depends on your contract. A client’s or an agency’s terms can say who owns the memory made on their jobs, and whether their texts may help you on anyone else’s. Read the terms you agreed to, and ask a lawyer about a specific case. Keeping one memory for each client makes it easy to give a client their own memory back, or to leave it out of work for others.