Translation memory (TM) is a database of previously translated source-language segments paired with their approved target-language translations. When a translator encounters the same or similar text again, a computer-assisted translation (CAT) tool can retrieve that previous translation for review and reuse. Rather than translating every recurring sentence from scratch, the translator can focus on new or changed material.
For organisations translating product manuals, software strings, contracts or regularly updated documents, a well-maintained TM can improve consistency and reduce repeat effort. It does not guarantee a discount, eliminate the need for review or generate new translations by itself. Savings depend on usable matches, the agreed pricing model, language pair, content risk and whether previous translations are still correct.
This guide explains how translation memory works, exact and fuzzy matches, a transparent hypothetical cost calculation, differences from machine translation, and the practical decisions buyers should make about quality, ownership and reuse. For a project quote, explore translation services or contact the translation team.
What is translation memory?
A translation memory is usually a collection of translation units. Each unit contains a source-language segment (often a sentence or software string) and its corresponding target-language translation. The system may also store language pair, client or project metadata, context, dates and review status.
For example, a company may already have an approved French version of “Save your changes.” The next time that English sentence appears in a suitable project, the TM suggests the stored French translation. The translator checks whether the UI context, audience, terminology and formality are still appropriate before accepting it.
Important: A translation memory is not a dictionary of individual words, although short terms may be stored as segments. A terminology database or termbase normally manages approved terms, definitions and prohibited alternatives. TM is also different from machine translation, which generates candidate text rather than simply looking up an approved existing bilingual segment.
How does translation memory work? A step-by-step example
- Import a source document. The CAT tool extracts translatable text from supported formats such as Word, XML or localisation files while retaining relevant formatting or tags.
- Split the content into segments. A sentence or interface string is compared against the project’s permitted translation memories.
- Retrieve matching translations. The system identifies exact matches, fuzzy matches or repetitions and displays the relevant match scores.
- Check the suggestion in context. A translator reviews accuracy, numbers, tags, tone, terminology and whether the stored translation is suitable for this occurrence.
- Translate and review new or changed segments. Anything without an appropriate existing match still needs translation and the agreed quality controls.
- Update the memory carefully. Approved translations can be saved with relevant metadata for future projects, according to the agreed reuse and ownership rules.
The process works best when documents contain stable, repeatable wording and the translation memory is well curated. If the source document changes substantially, many segments may be classified as new or fuzzy, reducing the practical benefit.
Exact matches, fuzzy matches, repetitions and context matches
The percentage shown in CAT software is a similarity indicator, not a translation accuracy score. A 100% match means the source segment matches an entry under that tool’s rules; it does not certify that the target translation is medically, legally or contextually correct.
| Input segment | Example result | Appropriate next action |
|---|---|---|
| “Save your changes.” | 100% exact match | Retrieve the approved previous translation; confirm the current UI context and terminology. |
| “Save all your changes.” | Fuzzy match (similar wording) | Use the previous translation as a starting point; revise and verify the changed meaning. |
| “Save your changes.” appearing again in the same project | Internal repetition | Check that identical wording is appropriate in both locations; apply the agreed repetition policy. |
| “Save your changes.” in an identical neighbouring or tagged context | Context match, sometimes displayed as 101% | Confirm the matching context and reuse as permitted by the project’s review rules. |
| “Configure the new dashboard.” | No useful match | Translate the new sentence and add the approved translation to the appropriate memory. |
Some tools provide a 101% context-match indicator when the source segment and its stored context match; certain systems also support an additional double-context category. Match bands and thresholds vary by product and settings. See memoQ documentation on translation memory match rates for an example of exact, fuzzy and context matching.
Why an exact match still needs judgment
Imagine the short English string “Open”. On a website it could be a button instruction; in a specification it may describe a status; elsewhere it could be an adjective. The same source characters do not always require the same target-language wording. The translator may need interface screenshots, context identifiers or nearby sentences.
Medical, legal and safety-related material needs particularly careful checking of outdated or ambiguous memory entries, quantities, warnings, changed laws and the target audience. An old error can be propagated very efficiently by a translation memory unless someone corrects the stored unit.
Translation memory vs machine translation, glossaries and TMS
| Technology | Primary role | Typical output |
|---|---|---|
| Translation memory (TM) | Retrieve prior source–target segment pairs | Previous human-approved or otherwise stored translation candidates |
| Machine translation (MT) | Generate a translation for source text using a model | New translated candidate, which may need specialist review |
| Terminology database / termbase | Manage approved words, terms and definitions | Term-level guidance and preferred or forbidden variants |
| CAT tool | Provide the editor where translators use TM, terms, QA checks and sometimes MT | A controlled workspace for creating and reviewing translations |
| Translation management system (TMS) | Coordinate files, assignments, approvals, languages, deadlines and suppliers | Project workflows, status tracking and access to language resources |
The tools can be combined. A CAT workspace may show TM suggestions, terminology alerts and MT candidates side by side, while a TMS routes the project between linguists and reviewers. For a wider comparison, read our translation software comparison for 2026 and introduction to translation management systems.
How translation memory can save time and money
Before quoting a repeat translation project, a supplier may run a TM match analysis that allocates words across new content, fuzzy matches, exact matches, context matches and repetitions. Different categories can require different amounts of work. The quote may use weighted words or agreed rates per category rather than billing all source words identically.
Illustrative pricing example — entirely hypothetical, not TranscribeLingo’s rates or a guaranteed discount: Suppose an English-to-French update contains 10,000 source words and the supplier and client negotiate the following category prices.
| Segment category | Example word count | Fictional agreed rate | Illustrative subtotal |
|---|---|---|---|
| New/no useful match | 4,000 | £0.120 per word | £480 |
| Fuzzy matches | 3,000 | £0.075 per word | £225 |
| 100% exact matches | 2,000 | £0.030 per word | £60 |
| Internal repetitions | 1,000 | £0.015 per word | £15 |
| Total | 10,000 | Variable by category | £780 |
If every word in this imaginary project had instead been charged at the example new-word rate of £0.120, the total would be £1,200. The example TM-weighted calculation is £780, or £420 (35%) less. This is only arithmetic designed to show the mechanism. Actual suppliers may use very different bands, minimum fees, review charges, specialist rates, project-management charges or no separate TM discount.
A memory full of outdated segments can also increase review effort. Quotes should therefore specify whether exact matches are reviewed, who pays for memory cleanup, how repetitions are counted and whether fuzzy matches need substantial retranslation.
When translation memory helps most
Video games also contain repeated interface messages, quests and terminology. Our guide to game translation and localisation shows how developers use stable string IDs, character context and release QA alongside translation memories.
| Content | Why reuse can help | What still needs review |
|---|---|---|
| Technical manuals and safety documentation | Repeated instructions, components and controlled phrasing | Warnings, model identifiers, measurements and changed requirements |
| Software and app localisation | Recurring menus, system messages and interface terminology | Variables, placeholders, character limits and interface context |
| Product catalogues and e-commerce listings | Repeated attribute descriptions, product categories and recurring text | Dimensions, product variants, marketing nuances and local terminology |
| Policies and recurring company documents | Stable clauses and regularly updated versions | Changes in obligations, dates, jurisdiction or business-specific meaning |
| Multilingual market research materials | Recurring survey questions or interviewer instructions | Cultural appropriateness, respondent wording and research methodology |
Translation memory is often less useful for a wholly new slogan, literary content or culturally adapted campaign copy that requires creative rewriting. For marketing material, a reused literal translation may harm rather than help a campaign if the audience and purpose have changed.
For organisations managing repeat documents, our business translation services page provides a commercial starting point; for source quality, see how to write source text for translation.
Translation memory quality: avoid spreading old mistakes
A shared memory can improve collaboration, but it must have practical governance. Without it, a translator may see contradictory translations for the same source sentence, an obsolete product name or wording originally approved for another client. Set a review process before allowing an imported memory to influence a new project.
- Approved entries: Decide whether new, draft or unchecked translations are allowed into the primary TM.
- Source and context: Preserve meaningful identifiers, surrounding text or document references where supported.
- Terminology changes: Update or suppress entries when a brand, regulation or product specification changes.
- Language variants: Keep British and American English, European and Latin American Spanish, and other relevant locales distinct when required.
- Duplicates and conflicts: Audit competing translations for the same source segment rather than blindly choosing the first suggestion.
- Project boundaries: Avoid making one customer’s confidential language assets available to an unrelated customer.
The translator and client should agree whether a 100% match can be automatically inserted, which risk categories always need checking and whether high-risk documents require independent review. Maintaining clean translation memories can make future projects faster without sacrificing the ability to trace decisions.
Who owns the translation memory? Privacy and export
Ownership and usage rights depend on the contract, not merely on which software is used. A client may want a project-specific TM as a deliverable; a supplier may maintain its own general resource. Do not assume the client automatically owns every shared memory, or that the agency can freely reuse confidential client content. Clarify rights before work begins.
- Ownership and delivery: Define whether the client will receive a project TM or only the translated files.
- Export format: Ask whether the memory can be supplied in an interoperable format such as TMX (Translation Memory eXchange).
- Permitted reuse: State whether entries are isolated by customer, subject matter or language pair, and what uses are prohibited.
- Confidentiality and data protection: Consider personal information, proprietary content, authorised storage locations, access controls and any data-processing terms.
- Portability and termination: Agree on export, handover, deletion and any limitations on migrating memories to another tool or supplier.
TMX is widely supported for transferring stored bilingual translation units between compatible tools, although metadata, tags and segmentation may require checking after import. For an example, memoQ documents TMX import and export. Before switching suppliers, test a sample export rather than presuming complete round-trip equivalence.
Choosing a TM-enabled translation workflow
- Identify repeat content. Estimate which documents, language pairs and product versions have recurring text.
- Assess existing language assets. Check previous translations, glossaries, bilingual files and any available TMX exports for quality and permissions.
- Agree segmentation and match rules. Decide on exact/fuzzy thresholds, context treatment, repetitions and whether all matches require review.
- Define terminology and quality controls. Set glossaries, source-reference materials, editing rules and reviewer responsibilities.
- Clarify pricing. Request a match analysis, explain the discount bands and include any minimum charges or cleanup work.
- Protect language assets. Specify ownership, access, retention, client separation and export obligations.
- Trial a real update. Compare effort, consistency and final quality before assuming large-scale savings.
A CAT tool may be useful for a freelance translator; a larger programme may also need a TMS. The relevant question for a translation buyer is whether the workflow improves the accuracy, governance and economics of their recurring content—not whether the supplier happens to advertise the newest platform.
Frequently asked questions
Is translation memory the same as machine translation?
No. Translation memory retrieves previously stored bilingual segments. Machine translation generates a candidate translation. Modern workspaces can combine both, but their outputs have different provenance and review considerations.
What is a 100% TM match?
An exact match is a source segment that matches a stored source segment under the tool’s matching rules. The stored target translation still needs to be appropriate for the new context, terminology and intended use.
What does a fuzzy match mean?
It is a stored segment that is similar, but not identical, to the new source. The translator must identify the difference and revise the candidate accordingly. The similarity percentage is not the percentage of words translated correctly.
Does translation memory always reduce translation prices?
No. A discount depends on real match counts, review effort, the source document, the contract and the supplier’s rate structure. Specialist quality checks, memory cleanup and project minimums can affect the final quote.
Can translation memory be used for medical and legal translations?
It can assist with recurring terminology and approved clauses, but a reused match is not automatic evidence of legal accuracy, clinical safety or regulatory acceptance. Translators must review updated requirements, context and critical content.
Can I transfer my translation memory to a new translation provider?
Potentially, if your contract gives you the required rights and the memory can be exported in a compatible format such as TMX. Verify content, metadata, locale, permissions and import quality before migrating.
Is translation memory useful for small projects?
Sometimes. A short revision of an earlier document may contain many reusable segments, while a much larger creative document may contain very few. The content’s repetition and quality of the existing memory are more important than file length alone.
Need translation for recurring business content?
If you have updated manuals, product catalogues, technical documents, policies or multilingual releases, send the source files, language pairs, deadlines and any approved previous translations. We can discuss project requirements and whether a TM-assisted workflow is relevant; no discount or turnaround should be assumed before assessment. Explore business translation services, professional translation options or request a quote.
