Translation Memory: how it works
A Translation Memory (TM) is a database of every sentence-level translation your team has ever produced. Every time Taia's AI or a human linguist translates a segment, that segment pair (source + target) is saved. The next time a similar sentence appears, Taia retrieves the stored translation and suggests it automatically.
The core idea
Imagine you've translated your software UI into French. Six months later, you update the product and need to translate it again. Around 70% of the sentences are identical or nearly identical to last time. Without a TM, you'd pay to translate them again. With a TM, those 70% are applied automatically — you only pay full rate for the genuinely new content.
Match scores
Not every stored segment is a perfect match. Taia assigns a match score (0–100%) based on how similar the stored source text is to the current segment:
- 100% — Identical. The stored translation is applied as-is.
- 95–99% — Near-exact (punctuation or casing differs). Apply with a quick review.
- 75–94% — Fuzzy match. The words are similar but not identical. Use as a starting point.
- Below 75% — Too different; not suggested.
For Professional translation orders, higher match scores mean a lower per-word rate. See Understand match scores.
How TM fills up
A TM grows from two sources:
- Active translation — every segment a linguist confirms in the CAT editor is saved into the TM automatically.
- TMX import — you can upload existing translation memories from other tools in TMX format. See Import a TMX file.
TM and AI translation
When a TM is attached to a project, Taia's AI engine consults it before translating each segment. A 100% TM match is applied directly — the AI doesn't re-translate it. For fuzzy matches, the AI uses the stored translation as a reference, which improves consistency even in new content.
This is how your TM trains the AI to sound like your brand over time: the more your team has translated in a particular style, the more the AI draws on that context for future projects.
Languages and regional variants
Taia never takes a translation from a different variant of your project's language. That holds for matches applied to your files, the match analysis on a quote, the suggestions in the CAT editor, and the examples the AI translation draws on. Each regional or script variant counts as a language of its own:
- A
de-AT(Austrian German) translation is never used for ade-DE(German, Germany) project, and the other way round. - A TM entry stored with a plain code such as
deorES(common in TMX files from other tools) doesn't say which variant it is. Taia can still show it as a suggestion in the CAT editor and give it to the AI as an example, ranked below an entry in your exact variant that matches just as well. It's never applied to your file automatically, and it never counts towards the match analysis on a quote. For languages where the variant is the script, such as Serbian and Chinese, a plain code is never used, because it doesn't say which alphabet the text is in. sr-Latn(Serbian, Latin script) andsr-Cyrl(Serbian, Cyrillic script) never match each other. The same goes forzh-CN(Simplified Chinese) andzh-TW(Traditional Chinese).
Upper or lower case in a language code makes no difference. A TMX file from another tool that labels its German DE-DE matches a de-DE project exactly.
So if your existing memory is British English and your new project is American English, the memory won't supply matches for it. To reuse it, create the project in the same variant as the memory.
Multiple TMs
You can create multiple TMs — one per language pair, one per product, one per department — and attach multiple TMs to a single project. Taia queries all of them and uses the highest-scoring match.
A TM can be thrown away
A memory that has learned the wrong things is not a permanent problem. Any TM can be deleted and replaced with an empty one at any time, and doing so leaves your projects and delivered translations untouched.
This matters most while you are evaluating Taia. Every edit a reviewer makes goes into the memory, test edits included, so a memory built during a trial is a mix of real terminology and experimentation. Practise freely, then delete the TM and create a clean one before real work starts. See start a TM over.