Language assets
Translation memory
Save approved source and target segments, reuse exact matches, and review fuzzy matches when similar text appears.
A translation memory (TM) stores source text and reviewed translations. When TM is enabled, automatic exact reuse requires an approved, unambiguous target. Different approved translations of the same source remain alternatives for review. Similar segments appear as fuzzy suggestions in the editor. A text match does not guarantee the same context. Part of the Translation Memory workflow (Translator Pro and up).
How your TM fills up
- Approving segments in either editor stores the pair. This is the main path, and it means your TM is made of translations a human signed off on.
- Importing TMX files brings in memories exported by common CAT tools. You can also export the memory to TMX for use elsewhere.
- Entries carry their language pair, domain, and source document, so lookups stay relevant.
Update an entry or keep an alternative
Each memory entry can be edited. Update this entry replaces that entry’s target and metadata. Save as alternative creates another target for the same source while preserving the original. Approval save-back also preserves different targets. Automatic reuse does not choose arbitrarily between conflicting approved translations.
Auto-propagate repeats in Settings is optional and off by default for accounts without a saved preference. When enabled, approval can fill identical unapproved segments in the Translation Editor as drafts. Check each occurrence: the same source wording can require a different translation in context.
Review duplicates across a memory
Review TM library works on one selected memory. Find duplicates searches the selected library rather than just the current list page. You compare targets, statuses and notes and keep the entry you want. Only identical copies are removed; different translations or metadata remain. A full scan can take longer than editor suggestions on a large library.
Build memory from an existing bilingual document
Step by step with screenshots: Build a TM from an old translation.
- Alignment works on the original and translated Word files, or on pasted text with one paragraph per line, into a destination memory you choose.
- Suggest alignment uses local multilingual matching. It can pair one paragraph with two, leaves weak matches unmatched for manual review and keeps the original passage order, so reordered sections may need manual pairing. Pair by position works without the local model. No translations are generated.
- Only pairs marked Reviewed are imported. Editing a pair clears its check; repeating the import skips existing identical pairs.
Word imports include body paragraphs and table cells, excluding headers, footnotes, drawings and text boxes. Use sections of up to 200 nonempty paragraphs per side and files up to 50 MB. Paragraphs may contain up to 2,000 characters for matching; the local matcher also has a 256-token passage limit. Split longer passages before matching. Similarity scores are not translation-quality scores. Importing reviewed pairs does not use translation words.
How documents use your memory
Exact matches first
Before anything goes to the AI, every segment is compared with your memory. An exact match (the same words, ignoring letter case and formatting) is reused as saved and costs no words. If the same source was saved with two different approved translations, neither is reused automatically.
From the memories chosen for the document
The job reads the memories selected in Editor resources for that document, or all of your own memories when you have not chosen. Memories shared by your organization are read only when you selected them there. The editor suggestions and the Analysis report use the same selection.
Glossary priority
When your glossary outranks memory, an exact match that leaves out a required glossary term is skipped. See Glossary.
Everything else
With AI translation, the remaining segments are translated by AI. In a document you translate in the editor they stay empty; partial matches appear as suggestions beside each segment.
Match types in the editor
| Match | Meaning |
|---|---|
| 100% | The normalized source text matches an approved memory entry. Preview the translation, then explicitly accept it. The optional auto-apply setting fills empty units in the Translation Editor. |
| Fuzzy | A similar source meets your minimum match setting. Review source differences, the stored translation and any number or negation warnings before accepting. |
| New | Nothing similar in memory. With AI translation the model translates it; in a document you translate in the editor it stays empty until you write it or click Translate. |
Managing your TM
Language assets → Translation memory lists saved segments with search and filters. Edit or delete an entry when its target text is outdated or wrong. Concordance search searches the same data from inside either editor.
In CAT settings, auto-apply 100% matches fills empty Translation Editor units with exact matches. Auto-propagate repeats fills identical unapproved, unedited units when you approve a translation. Existing manual edits and protected PDF units remain unchanged; propagated units stay drafts for review.
Multiple memories and termbases (teams)
Business and Enterprise plans can keep several named translation memories and termbases. Containers are created and renamed in Language assets; one is active for new approvals, and existing segments can be assigned by language pair or domain. The same controls are available for termbases.
Named memories organize entries and identify their origin. Editor resources sets and orders the memories searched for each document. You can select no memories to turn library suggestions off. The memory used to save approved work is a separate choice. Eligible organization members can also select team resources; their sharing filters still apply.
When memory changes the job
An exact match can use the target text that was approved earlier instead of generating another version. A fuzzy match is offered as a reference and still needs review. Run the Analysis report before a job to see how much source text falls into each match band.
Keep exploring
Analysis report
See the source-word total, internal repetitions, translation-memory coverage, and a transparent workload estimate before you begin a job.
Concordance search
Search Translation Memory for a word or phrase and review the source and target segments where it appeared.
Translation editor
Review source and target segments side by side with Translation Memory, terminology, QA, comments, status, and per-segment retranslation.
Glossary
Store approved term pairs for a language pair and apply them during text translation, document translation, and review.
Revision comparison
Compare two versions of a document, see every class of change, and if the original was translated here, update only the sentences the revision touched.