Is translation memory the same as machine translation?

No. A memory retrieves stored source-target pairs; a translation system generates translations, although both can work together.

Can it help with completely new content?

It may retrieve similar fragments or segments, but a memory alone cannot supply every unseen translation.

Does a 100 percent match guarantee a perfect translation?

No. It generally indicates source-text matching under the tool's rules, not universal contextual correctness.

AI Glossary / Natural language and speech / Intermediate

Translation memory

Definition

A translation memory is a database of previously translated text segments paired with their corresponding translations. A computer-assisted translation tool retrieves these pairs when a new document contains identical or similar source text. The stored pairs are often called translation units. Translation memory supports reuse and consistency; it does not, by itself, generate a new translation for every possible sentence.

What it means

A translation memory preserves a relationship between source-language content and target-language content. A translator can inspect a suggested previous translation, reuse it, or adapt it to the new context. Repeated product instructions, support articles, and document revisions are common settings where this can be useful.
It differs from a terminology database, which generally records preferred translations and information for terms. It also differs from machine translation, which produces translations through a computational translation system. A modern workflow can combine all three, but their outputs and sources should remain distinguishable.

How it works

The tool segments incoming text and searches the memory for matching source segments. An identical source segment can receive an exact-match classification; a similar segment can produce a fuzzy match requiring adaptation. The translator reviews the target text and saves an approved translation according to the workflow.
Metadata can help distinguish project, language variant, context, and approval status. An exact source-text match is not necessarily a context match or an assurance that the stored translation remains suitable. Similarity scores describe the tool's matching process and should not be read as universal translation-quality percentages.

Three practical examples

A software team can reuse reviewed interface instructions when preparing a new release, while checking changed feature names and behavior.
A manufacturer can retrieve earlier translations of repeated maintenance passages and review revised quantities or warnings carefully.
A multilingual publishing team can use a shared, curated memory to reduce avoidable variation across recurring documents.

Think of it this way

Think of an organized notebook containing earlier sentences and their approved translations. When a similar sentence appears, the translator can consult the notebook instead of starting from nothing. The analogy stops at automatic correctness: even a well-kept notebook can contain old terminology, a different audience's wording, or a translation that is unsuitable in the current context.

Worked example

Consider an original miniature English-to-Spanish memory. It stores source segment Open the file. with target Abra el archivo. A new document contains that same source segment, so the tool can retrieve the stored target as an exact source-text match.
A second segment says Open the folder. It differs by one word and may retrieve the first unit as a fuzzy candidate. A translator must change archivo to carpeta and confirm that Abra la carpeta. fits the intended interface and audience. Reusing the old target unchanged would preserve the wrong object even though much of the sentence matches.
A third segment says Do not open the file. Its overlap with the stored sentence is large, but its meaning reverses the instruction. A suitable illustrative translation is No abra el archivo. This shows why a high textual similarity must not be treated as permission to copy the target without reviewing the difference. No specific fuzzy percentage is claimed, because scoring depends on the tool and configuration.
Now consider a separate fictional workload of 100 segments. Assume 40 exact matches each require 15 seconds of review, 30 fuzzy matches each require 45 seconds of editing and review, and 30 new segments each require 90 seconds. The total is 600 + 1,350 + 2,700 = 4,650 seconds, or 77.5 minutes. If all 100 instead required 90 seconds, the baseline would be 150 minutes. Under these stipulated assumptions, the difference is 72.5 minutes.
These figures illustrate workload arithmetic, not a promised productivity gain. They omit setup and final document review, and exact matches still consume time. If outdated entries require extensive correction, the actual benefit can shrink or disappear. Quality and elapsed effort must therefore be measured in the real workflow.

Illustrative case study

Fictional case: a small software publisher combines translation memories from several product lines. Reviewers notice that an identical source label has different intended meanings across products, and the merged database suggests inconsistent targets.
The team adds product context, distinguishes approved entries from unreviewed imports, and resolves conflicting terminology. Translators can see where a suggestion came from and can reject it when the current screen or document requires different wording.

Common misconceptions and mistakes

Do not assume an exact match means the translation needs no review. Context, product changes, language variant, and previous errors can all affect suitability. Do not equate a fuzzy-match score with the probability that a translation is correct.
A memory is not inherently limited to flawless human translations: imported or automatically generated text can enter a poorly controlled database. Its reliability depends on provenance and review. A larger memory is not always more useful if it mixes conflicting domains or accumulates obsolete entries.

Business use

Translation memory can reduce repeated effort and support consistent language when teams maintain it as a reusable content asset. The benefit is strongest when the collection contains relevant, well-reviewed material that recurs in future work.
Track who can add or approve entries, preserve useful context, and retire obsolete wording. Measure accepted suggestions, correction effort, and final quality rather than only match coverage. Keep private or client-specific content within appropriate access boundaries. When integrating AI translation, distinguish generated suggestions from approved memory entries so the review process remains understandable.

Check your understanding

Why is Do not open the file. a risky candidate for blindly reusing the translation of Open the file.?

Answer: The added negation reverses the instruction. Textual overlap is high, but the target must reflect the changed meaning.

Related terms

Translation unit
Machine translation
Localization
Parallel corpus
Alignment
Computer-assisted translation
Fuzzy matching
Terminology management
Human review

Learn next

Revise the workload example so exact-match review takes 30 seconds instead of 15. Calculate the new total, then identify which assumptions a real team would need to measure before using the estimate to plan a project. Separately check how a change in target-language audience might affect reuse.

Exact match

Sources

Primary product explanation: RWS Trados, What is a translation memory?: https://www.trados.com/learning/topic/translation-memory/ . Original resource paper: Steinberger and colleagues, DGT-TM: https://arxiv.org/abs/1309.5226 . Translation illustrations and workload calculations are original; they are not claims of measured product performance.

Previous: Translation-invariant prior Next: Translation unit