The basic idea
A translation memory is a database of text that has already been translated. It stores pairs: a segment in the source language alongside its approved translation. When a translator later works on new material, the tool compares each new segment against the stored ones. If it finds the same or a similar segment, it offers the earlier translation as a suggestion.
The idea comes from written translation, where documents often repeat sentences or change only slightly between versions. It has also found a place in subtitling, where series, training courses and product videos often repeat phrases.
Note that a translation memory is different from machine translation. A memory stores human-approved translations of specific segments and reuses them. Machine translation generates new text. Many workflows combine the two, offering a stored match when one exists and a machine suggestion when it does not.
How matches work
When a memory tool compares segments, it usually reports a degree of similarity:
- Exact match. The new segment is identical to a stored one. The stored translation can often be reused directly, though the translator should still check the context.
- Fuzzy match. The new segment is similar but not identical. The tool offers the earlier translation, with differences marked, and the translator adjusts it.
- No match. Nothing similar is stored, so the translator works from scratch.
The thresholds that tools use to call something a fuzzy match differ, and they can often be configured. A match that looks close on paper may still be wrong in context, which is why a person remains responsible for each decision.
Why it helps in subtitling
- Consistency. If a recurring phrase, title, greeting, warning or instruction was translated one way before, the memory prompts the same wording again. Viewers of a series or course see the same terms throughout.
- Terminology. Together with a term glossary, a memory helps keep names and technical words stable across episodes, modules and translators.
- Speed on repeated content. Training and corporate videos often reuse standard text such as intros, outros, legal lines and calls to action. Reusing those saves effort.
- Updates. When a video is revised and only parts change, a memory helps carry over the unchanged translation. This is useful after a video has been re-edited.
- Knowledge that outlasts individuals. A well-kept memory lets new translators pick up the established style, rather than starting afresh.
Why subtitles are awkward for memory tools
Subtitling differs from document translation, and some of the differences reduce the benefit.
- Segments are short and fragmentary. A subtitle may be half a sentence, split to fit a line or a timing. Two identical fragments may belong to different sentences with different meanings. Matching on the fragment can mislead.
- Context is crucial. The right translation of “Right,” “Well,” or “You too” depends on the scene. A stored match may be right in one place and wrong in another.
- Splits differ between languages. The way a sentence is divided into cues reflects the source language’s rhythm and the timing, not the target’s grammar. Memory based on cue boundaries may capture awkward breaks.
- Length constraints. An earlier translation may have been condensed for a particular cue and be too long or too short for a new one.
- Timing is separate. The memory stores text, not timecodes or reading speed. Reused text still has to be fitted into its new cue.
- Spoken language varies. The same idea is said in many ways, so exact matches may be less frequent in dialogue than in manuals.
Some teams address this by storing whole sentences rather than individual cues, or by making sure the text is aligned sentence by sentence before it goes into the memory.
Keeping a memory healthy
A memory is only as good as what goes into it. Poor entries get reused, and errors repeat. Good practice includes:
- Add only reviewed, approved translations. Do not store drafts or unchecked machine output.
- Separate memories by project, client, subject or language variety. Mixing, for example, two different style preferences causes inconsistency.
- Clean up regularly. Remove outdated terms, duplicates and entries that conflict with each other.
- Record corrections. When a reviewer fixes a translation, update the memory so the mistake is not offered again.
- Keep notes with the memory. Where a stored match applies only in a certain context, note it.
- Watch for confidentiality. A memory may contain sensitive or proprietary text. Think about who can access it and how it is shared with outside providers.
- Check ownership expectations. If you work with an outside provider, it is reasonable to clarify who holds the memory built during the project and whether it can be exported in a standard form.
Using memory with other tools
A memory fits alongside other quality aids: a style guide for decisions, a glossary for terms, an automatic checker for line length and reading speed, and a human review. None replaces the others. A memory can encourage consistency, but it cannot judge whether a stored translation is right for this scene.
When it is worth it
Memory pays off most for projects with repetition, long-running series, multiple languages, several translators or frequent updates. For a single short video with no repeated material, the setup effort may not be worth it. For large multilingual programs, it can be a significant part of the workflow.
A sensible expectation
Think of the memory as a reliable colleague with a good recall of earlier work, but no view of the current scene. It offers suggestions, and the translator decides. Treat exact matches with a quick check, fuzzy matches with real attention, and keep reviewing the memory itself. Used that way, it supports consistency without taking over judgment.