Translation Memory Explained: How Localization Teams Reuse Approved Translations

What a translation memory stores

Each entry in a translation memory is a pair: a source segment, usually one sentence, and the translation a human approved for it. Most entries also carry metadata such as the date, the translator, the client, the project and sometimes a quality status. Over years, a company's memory can grow to hold every sentence it has ever had translated.

The memory is built as a side effect of normal work. Professional translators use computer-assisted translation tools, usually called CAT tools, which split a document into segments, show each one next to a box for the translation, and save every confirmed translation into the memory automatically.

It is worth being clear about what a TM is not. It isn't machine translation. A TM never generates new language; it recalls translations people already approved. That is exactly why localization teams value it: consistency and a known quality history.

Segments, exact matches and fuzzy matches

When a new document is opened in a CAT tool, it is segmented, and each segment is compared with the memory. The result for each segment falls into one of a few categories:

Exact match

The same sentence was translated before; the stored translation is offered as is

Context match

An exact match whose surrounding sentences also match, so it is even more likely to be right

Fuzzy match

A similar sentence, scored by how much it differs; the translator edits the stored version

No match

Nothing similar; the sentence is translated from scratch or from a machine suggestion

Repetition

A sentence that appears more than once in the same new document

A worked example: last year's training script contained "Press the red button to stop the machine." This year's says "Press the green button to stop the machine." The CAT tool finds a high fuzzy match, shows the old translation with the difference highlighted, and the translator only changes the color.

Agencies commonly price work by these categories, charging less for exact and fuzzy matches than for new sentences, because they take less effort. If a localization quote includes a match analysis, that is what it is counting. The details and thresholds vary between vendors, so ask how yours calculates it.

Termbases: terms rather than sentences

A translation memory works on whole sentences. A termbase works on individual terms. Each termbase entry records a concept, its approved term in each language, often a definition, and sometimes forbidden alternatives. While a translator works, the CAT tool highlights terms in the source segment and shows the approved translation, and some tools flag a segment where the approved term wasn't used.

Translation memory

Whole segments; reuses past translations; grows automatically

Termbase

Single terms and names; enforces vocabulary; curated by hand

Style guide

Rules for tone, register, numbers and formatting; written by hand

The three work together. The termbase keeps "Pro plan" as "Pro plan" in every language; the TM reuses last year's approved sentence about it; the style guide says whether to address the viewer formally. Building the term list is covered in translating names and technical terms, and the style guide in creating a translation style guide.

Standard file formats

Translation memories and termbases are meant to outlive any single tool. Two exchange standards are widely supported: TMX, the Translation Memory eXchange format for memories, and TBX, the TermBase eXchange format for terminology. Both are XML. XLIFF is a related standard for exchanging the content being translated. Support and versions differ between tools, so check the documentation of the tool you use.

The practical point for a buyer: if an agency builds a memory from your content, ask whether you can receive it as a TMX file. That keeps your translation history portable if you change vendors.

How translation memory and machine translation work together

In most professional workflows today, the two are layered. For each segment, the CAT tool offers a TM match first if a good one exists. Segments without a good match get a machine translation suggestion, which the translator post-edits. Every approved segment then goes back into the memory, so the memory improves the next job, regardless of where the first draft came from.

Some machine translation systems can also be adapted using a company's memory, so their suggestions follow its terminology and style more closely. The editing step in this workflow is covered in machine translation post-editing.

Why video gets less out of a TM

Translation memory was designed for written text that repeats: manuals, help centers, software strings, legal boilerplate. Video is harder for several reasons.

• Speech rarely repeats word for word. Two videos about the same feature will describe it in different sentences, so exact matches are rare.
• Subtitle cues aren't sentences. A sentence split across two cues becomes two odd segments, and a cue containing the end of one sentence and the start of another matches nothing well.
• Spoken language is messy. Fillers, restarts and fragments produce segments that are unique by accident.
• Dubbing scripts change for timing. A translation adjusted to fit a slot is right for that video but may not be the best entry to reuse elsewhere.

Where a TM does help with video: recurring intros and outros, standard disclaimers and safety statements, product and feature names through the termbase, and training modules that are re-recorded with small changes each year.

Keeping a lightweight glossary and phrase list yourself

Without a CAT tool, you can still get most of the consistency benefit for video with a shared spreadsheet:

1. Create one row per term, with columns for the source term, each target language, a do-not-translate flag, notes and the date it was approved.
2. Add a second tab for recurring sentences: your channel intro, sign-off, disclaimers, calls to action. Store the approved translation of each.
3. Seed both tabs from the transcripts and translated transcripts of your first few reviewed videos.
4. Give the spreadsheet to every reviewer, and ask them to check each new translation against it using search.
5. When a reviewer approves a new term or phrase, add it the same day.
6. Review the list every few months and remove entries for products or campaigns that no longer exist.

This is a manual translation memory and termbase, and for a small video library it is often all you need.

Limits and trade-offs of translation memory

• A memory reuses mistakes as faithfully as good translations. An error approved once can spread into dozens of later documents unless someone cleans the memory.
• Exact matches can still be wrong in a new context, for example when the same short sentence refers to different things.
• Memories need maintenance: outdated product names, old style decisions and duplicate entries accumulate.
• CAT tools and memory management take time to learn and often cost money, which rarely pays off for occasional, one-off video translation.
• Ownership can be unclear. Agree in your contract who owns the memory built from your content.

Translation memory and mydubly

mydubly has no translation memory, termbase or glossary feature. Each job is translated fresh by its translation engine, without access to earlier translations of your other videos, and you can't upload a TMX file or term list. Even within one long video, the text is translated in pieces, which is why terminology can vary; the effect of that on long recordings is discussed in translating long videos.

What mydubly does give you are exportable texts that fit into a TM-based workflow: a timestamped transcript, subtitles as SRT and VTT, and an optional translated transcript. Many CAT tools can import subtitle files, so an agency or freelancer can bring your translated SRT into their tool, check it against your memory and termbase, and add the approved result to the memory. A translated transcript for a 10-minute video costs 10 credits (1¢). The video translator page describes the outputs, and the SRT generator page explains the subtitle format.

Next step

If you already work with a localization vendor, ask whether they keep a translation memory and termbase for your content and whether you can get them as TMX and a spreadsheet. If you don't, start the two-tab spreadsheet above with your next reviewed video. For choosing between vendors, freelancers and AI tools in the first place, see choosing a video localization vendor, and for the bigger picture, the video localization overview.