Translation memory
Reuse past translations — exact matches auto-fill for free, fuzzy ones are suggested.
6 min read
What the memory stores
Every translated block feeds a per-project, per-language memory: the source text, its translation, and whether it came from the machine or from a human edit. Human pairs are never downgraded by machine ones, and usage counts track what actually gets reused.
Automatic reuse
When a block’s source matches the memory exactly, the translation is filled in without calling the model — zero credits, instant. A page fully covered by the memory skips dispatching altogether. Fuzzy matches (down to 80% similarity) are proposed instead of applied.
Re-running OCR resets blocks — the memory is what keeps your past manual work reusable.
Suggestions in the editor
In the script panel, blank rows show TM chips — TM · exact or TM · 87%, with a human marker when the pair was curated by hand. Click to apply one, or apply every exact match on the page in bulk. Suggestions can be toggled in the editor settings.
- 1Open the script panel on a page with untranslated blocks
- 2Look for TM chips on blank rows
- 3Apply a single suggestion, or Apply N TM for all exact matches
- 4Your manual edits flow back into the memory automatically
Browse, edit, import & export
The project’s Translation memory page lists every pair with search, language, chapter and origin filters, sortable columns and bulk delete. Each pair shows the chapters it came from — so you always know why it exists. Import industry-standard TMX files (up to 20 MB) — imported pairs count as human-curated — and export each language’s memory as TMX for your other tools.
Deleting a chapter removes its pairs from the memory — unless another chapter or a TMX import still relies on them.
Consistency audit
The audit view groups identical sources that were translated differently across the project. Review the variants and promote one — it is pushed everywhere the source appears, including already-translated pages.
