Translation Memory: Who Owns It, What It Is Worth and How Manufacturers Keep Control of It

A translation memory belongs to whoever the contract says it belongs to, and in most translation relationships the contract says nothing. Settle ownership when a cooperation starts, not when you try to leave, and ask for the memory in TMX. At DEMA Solutions, the translation memory is always the client’s: exportable on request, with glossaries included, from the first project. When that moment comes, our step-by-step guide to switching translation supplier and keeping your memory shows how to request, check and consolidate the files.

Automated production line in a modern factory

Who lives with translation memory decisions after the order is signed?

This article is written for the people who live with the consequences of translation decisions long after the purchase order is signed: technical documentation managers, technical writers and localization managers, and the procurement, operations and quality directors who approve translation budgets at manufacturers of machinery, HVAC systems, lifting and handling equipment, packaging and process lines or automation. It is also for Italian companies that belong to international groups, where the parent wants the same terminology everywhere and the subsidiaries keep re-checking every translation they receive.

The question usually comes at a turning point: a new market or product line, documentation for the Machinery Regulation (EU) 2023/1230 that applies from 20 January 2027, supplier consolidation after an acquisition, or a change of supplier after inconsistent terminology or a unit error in a manual. Someone then asks: “We already have translation memories from our previous supplier. Can you work with those?” The sections below explain what that memory is, who owns it, what it is worth, and how to take control of it without disrupting work already running.

What a translation memory is, and what it actually stores

A translation memory is a database of source sentences paired with their approved translations. Every time a translator works on your content, each sentence (a “segment”) and its translation are saved. When new content arrives, the memory compares every new segment with everything already stored and offers identical or similar translations for reuse. Translation memory software does the matching; the value sits entirely in the content, which is your own documentation, translated and approved over years.

For a manufacturer, that content is very specific. It holds the safety instructions and warnings that must read the same in every manual, the component names, error messages and process terms shared between the operating manual and the machine’s HMI strings, the descriptions repeated across machine variants and spare parts catalogues, and the many small decisions about how your company says things in each language. A well-kept memory also stores useful context: which document a segment came from, when it was created or last changed, who approved it, and sometimes the product line or client division it belongs to.

That is why a translation memory is more than a cost-saving device. It is the record of your technical language in every market you sell to. Losing it does not just raise the price of the next project. It resets consistency to zero, and it is one of the main reasons technical manual translation costs climb from one release to the next.

A memory is not the same thing as a glossary, and you need both. The memory stores whole sentences; the glossary, or term base, stores individual approved terms with their definitions and forbidden alternatives. The memory saves time and money on repeated text, while the glossary stops the same component from being called three different things across a manual, a catalogue and a service bulletin. Our article on keeping technical terminology consistent across translators and suppliers explains how to build a glossary that people actually use.

Who owns a translation memory?

Because the memory is built from your source content and paid for through your translation invoices, most companies assume it is theirs. Legally and practically, it is less simple. If ownership is not written into the offer, framework agreement or contract, the supplier holds the only copy inside its own systems and may treat it as a working file. Some suppliers hand it over without discussion. Others send it late, send only part of it, send it without the glossary, or ask for a fee. A few have no clean way to export it at all, because years of work were stored in a format nobody else can read.

When a manufacturer comes to us after years with another provider, the most common situation is not a dispute but uncertainty. Nobody can say where the memory is, which languages it covers or whether it was ever kept separate from other clients’ content, which makes a clean handover harder and raises its own confidentiality questions. Ownership should therefore be stated in writing, together with the export format, timing, cost and what happens to the supplier’s copy at the end. Content you paid to translate, built from your own documentation, should never sit somewhere you do not control.

What to agree before the first project, and what to ask a supplier you already use

The cheapest moment to settle ownership is before anything has been translated. The memory starts growing with the first project, and every sentence approved from that day on becomes part of an asset you will want to keep. If you already work with a supplier, you do not need to change anything to protect your memory. You need clear answers to six questions, ideally in writing:

  1. Who owns the translation memory and glossary built from our content? The answer should be “you do”, without conditions.
  2. In which format can we receive them? Ask for TMX for the memory and TBX or a structured spreadsheet for the glossary. A proprietary backup or a PDF of past translations is not the same thing.
  3. Is our memory kept separate from other clients’ content? A dedicated memory per client, or per division if you have several, is far easier to hand over and to maintain.
  4. How quickly, and at what cost, can we get an export? “On request, at any time” is the answer to look for. An export fee or a long notice period is a warning sign.
  5. How will we see what the memory saves us on each project? You should receive an analysis showing the match bands for every job, not just a total price.
  6. What happens to the memory when the contract ends? It should be handed over in full, and the contract should say whether the supplier deletes or retains its copy.

If you are starting a cooperation, make these answers part of the agreement. If you already work with a supplier and any answer is vague, fix it now, while the relationship is working well. It is much harder to negotiate ownership when you are already on your way out. The same questions belong in any wider assessment of a supplier, alongside the points in our guide on how to choose a technical translation company.

Match types and what they mean for cost and review

Ownership matters because the memory drives what you pay. Translation of repeated content is usually priced by how closely each new segment matches the memory, and this is where most misunderstandings start, both in procurement and on the documentation team.

Exact and context matches. A 100% match means the same sentence has already been translated and approved. A context match goes one step further: the sentence is identical and so are the segments around it, so the translation almost certainly still fits. These cost least, but they are not automatically free. In safety-critical documents someone should still confirm that the approved sentence suits its new place, because a warning that was correct next to one illustration may be wrong next to another.

Repetitions. Sentences that appear more than once in the same job are translated once and reused. Spare parts catalogues and variant manuals are full of them, which is one reason structured documentation is cheaper to translate than it looks.

Fuzzy matches. A fuzzy match is a similar sentence: perhaps one value, one component name or one word order has changed. An 85% match does not mean the linguist does 15% of the work. Someone still has to find the difference, check whether the context has changed, update numbers, units and tags, and correct the grammar around the edit. For technical content that can take almost as much attention as a new translation. Low and mid-range fuzzy matches, roughly 50 to 75%, deserve a full professional check, especially when your glossary or product has changed since the original translation.

New words. Everything below the lowest fuzzy band is translated from scratch.

Two warnings apply to every pricing model. First, reused text is not risk-free: a memory can carry old errors and outdated terms forward, and a term changed in your glossary last year may still appear in older segments. Second, stacked discounts can hide real work. When machine translation is combined with memory matches, some pricing models apply one discount on top of another; the quote looks cheap, but the human work needed to make the text publishable has simply been underpaid, and it usually shows in the result. The same trap appears in machine translation post-editing quotes.

How to measure leverage and compare quotes fairly

Leverage is the share of a new job that the memory can supply, in full or in part. Procurement teams often compare suppliers on the per-word rate alone, which says very little about what a manual update will really cost. What matters is the effective cost of each job once the match bands are applied, and whether that figure is visible to you.

A simple illustration shows why. Imagine an update to an operating manual of 10,000 words. If 4,000 words are exact or context matches, 3,000 are fuzzy matches and 3,000 are new, the price depends on how each band is weighted, not on the headline rate. A supplier with a slightly higher rate but a well-maintained memory of your content will often cost less per update than a cheaper supplier starting with no memory, or with a memory so inconsistent that every match needs rework. The numbers in this example are illustrative only; your own analysis reports will show the real distribution for your documents.

To compare quotes fairly, run the same source files against the same memory and ask each supplier for its analysis by band. Look at the trend over time, too. On a stable product range, leverage should grow as the documentation history grows. If it stays flat or falls, the memory is probably fragmented, polluted with unapproved text, or not being used consistently. Clear reporting of this kind is part of running a translation workflow that does not depend on email and spreadsheets.

Formats and portability: TMX, TBX and what a handover should include

TMX (Translation Memory eXchange) is the open, XML-based format that professional translation memory software can import and export. A TMX file contains your source and target segments, the language codes and, when exported carefully, useful metadata such as creation dates, authors and custom fields. TBX (TermBase eXchange) does the same job for glossaries. Together, they are what makes your linguistic assets portable between suppliers and systems.

Portability is not only about the format name, though. A TMX file can be technically valid and still disappoint: inline tags from FrameMaker, InDesign or DITA topics may have been flattened, metadata may have been stripped, or segments from several clients may be mixed together. When you request an export, a complete handover usually includes:

  • One TMX file per language pair, or one multilingual file, covering all your content and nothing else
  • The glossary or term base in TBX or a structured spreadsheet, with forbidden terms and definitions
  • Metadata kept where possible: dates, project or product line, and approval status
  • A short note on any segments the supplier knows to be outdated or unapproved

Before you rely on the export, open it in a second environment and check a sample: are the languages right, do the tags survive, and does the segment count look plausible for the years of work it represents? This check takes an afternoon and saves a great deal of uncertainty later.

Fragmented memories across suppliers, subsidiaries and acquisitions

Many manufacturers do not have one translation memory. They have several: one per former supplier, one inside a subsidiary in Germany or Spain, one inherited with an acquired company, and a folder of old Word and PDF translations that never went into any memory at all. Each was built with different glossaries, different style choices and sometimes different source versions of the same manual.

The result is familiar: the same component translated three ways, subsidiaries re-checking every delivery, and leverage that never grows. In international groups it is sharper, because the parent wants group-consistent terminology while each local team has its own approved texts. Managing this across several suppliers and markets is the subject of our article on how organisations manage translation across multiple vendors and markets.

Consolidation means bringing these sources together into one clean memory per language pair, with a single approved glossary on top. Old translations that exist only as documents can be aligned with their source files to create new memory content, which is the work of translation memory alignment. Conflicting translations of the same sentence have to be resolved rather than simply merged, usually by checking them against the current glossary and asking the people who know the product. Done well, consolidation turns several partial assets into one that works for every supplier and every subsidiary.

Translation memory management: keeping the asset healthy

A memory left alone slowly degrades. Products change, terms are replaced, a supplier saves an unrevised draft, and a sentence that was correct for one machine generation is reused on the next. Translation memory management is the ongoing work that stops this from happening, and it does not have to be complicated.

Approve before you store. Only revised and approved translations should go into the master memory. Drafts, rejected versions and machine output that nobody has checked should never be saved as reusable text.

Keep the glossary and the memory in step. When a term changes, search the memory for the old term and decide, segment by segment, whether to update it or mark it. Otherwise every future match will quietly reintroduce the term you have just retired.

Remove duplicates and contradictions. Over the years, a memory collects several translations of the same sentence. Periodic cleaning removes exact duplicates and settles conflicting entries so that the memory offers one approved answer, not three.

Use metadata. Tagging segments by product line, document type or division lets you prioritise the right matches, so a sentence from your packaging machinery documentation is preferred over a similar one from an unrelated product.

Check reused text again where risk is high. For safety warnings, declarations of conformity and anything that goes into a CE technical file, a second look at reused segments is a small cost compared with an error carried into every language. How a supplier handles this is one of the points worth checking when you assess a supplier’s quality before you hire them.

Translation memory and machine translation

Machine translation has not made the translation memory obsolete. It has made it more important. Machine translation generates new text from statistical patterns; a translation memory reuses text your company has already approved. Used on its own, a generic engine knows nothing about your component names, your safety phrasing or the choices your subsidiaries have agreed to.

The best results come when machine translation is controlled by your own memory and glossary: approved segments are reused first, machine output is generated only for genuinely new text, and that output follows your terminology before a qualified linguist post-edits and a second linguist revises it. Every approved result then goes back into the memory, so the share of approved text keeps growing. This only works in your favour if the memory is yours. If it sits with a supplier, every improvement you pay for strengthens an asset you cannot take with you.

A step-by-step plan for taking control of your translation memory

You do not need a large project to get your linguistic assets under control. Most manufacturers can do it alongside normal translation work in a few steps.

Step 1: Make an inventory. List every supplier, subsidiary and internal team that has translated your documentation in recent years, with the languages and document types involved. Include the old translations that only exist as Word, InDesign or PDF files.

Step 2: Request the exports. Ask each current and former supplier for your memory in TMX and your glossary in TBX or a spreadsheet. Keep the request simple and factual; most suppliers will cooperate when asked clearly.

Step 3: Check what you received. Open the files, check language pairs, tags and segment counts, and note any gaps or obvious mixing with other clients’ content.

Step 4: Approve the glossary first. Agree the key terms with your engineers, service team and, in a group, the parent company’s documentation owners. The glossary is the reference against which conflicting memory entries will be resolved.

Step 5: Consolidate and clean. Merge the memories into one master per language pair, align any old document-only translations, remove duplicates and resolve contradictions against the approved glossary.

Step 6: Put ownership in writing. Update your framework agreements so that every supplier works on your master memory, delivers updates back into it and hands it over in full on request.

Step 7: Track leverage on every job. Ask for a match analysis with each quote and watch the trend over time. Rising leverage on a stable product range is the clearest sign that the asset is working.

How DEMA Solutions helps

When manufacturers use our technical translation services for manuals, catalogues, HMI strings and technical files, the memory and glossary built along the way belong to them from the first project and are handed over as TMX and TBX on request. Memories inherited from a previous supplier are taken as the starting point, not discarded, and you get a frank assessment of their condition.

Every translation is revised by a second qualified linguist under our ISO 17100 certified process, which our quality page describes. In sensitive projects, matches pulled from the memory are checked again instead of being trusted automatically. Our in-house DTP team works in the original InDesign, FrameMaker or XML files, so tags and layout survive into each language and the memory stays clean.

Behind this is a simple position: any supplier could work from the same memory, so we expect to win each project on the work itself, not because your linguistic history is locked in with us.

A practical case: an Italian maker of industrial equipment that exports most of its production ships each machine with manuals, safety information and interface texts in the buyer’s language, and many new models are variants of an existing base design. A memory dedicated to that client lets content repeated across variants be recognised and reused, while an approved glossary keeps component and process names aligned across manuals, interface texts and spare parts lists. Each new project benefits from the history built before it, as the industrial manufacturer case study explains.

Where TUBE fits

TUBE is built on the principle this article argues for. The memory that grows with every approved translation is yours, and you can export it whenever you want, to keep a copy, to audit it or to take it elsewhere. Because that memory and your glossary guide machine translation, its value shows up on each project, and a leverage report puts a figure on how much of every update it covered.

Frequently asked questions

It depends on the contract. Because a memory is built from the client's own source content, many companies assume it is theirs, but if ownership is not written down, a supplier may treat it as its own working file. State ownership, export format and handover terms explicitly in the agreement.

That the memory and glossary built from your content belong to you, that you can receive them in TMX and TBX (or a structured spreadsheet) at any time, at no cost or a stated cost, that your content is kept separate from other clients', and that everything is handed over in full when the cooperation ends.

Usually yes, if you ask clearly and the contract does not restrict it. Request the memory in TMX together with the glossary, then open the files in a second environment and check languages, tags and segment counts before you rely on them.

TMX (Translation Memory eXchange) is an open XML format for exchanging translation memories between systems. It stores source and target segments with their language codes and, when exported carefully, metadata such as dates and authors. It is the format to ask for whenever you want your memory to be portable.

Because a partial match still needs a professional to find what changed, adapt the sentence to its new context and check numbers, units, tags and terminology. Only exact matches and repetitions need little or no rework, and even those deserve a check in safety-critical text.

Collect TMX exports from every source, approve a master glossary first, then merge the memories into one per language pair, removing duplicates and resolving conflicting translations against the glossary. Old translations that exist only as documents can be aligned with their source files to recover them.

No. Machine translation generates new text, while a memory reuses text your company has already approved. The two work best together, with approved memory matches used first and machine output limited to new content, then post-edited and revised by qualified linguists.

Make sure your translation memory is yours

If you are about to start a translation cooperation, settle ownership before the first project. If you are not sure who holds your translation memory today, or what state it is in, that is the first thing to find out. Tell us what you translate and who has handled it so far, and we will help you recover, consolidate or set up a memory that stays under your control.

How can we help?

Tell us what you need and we’ll get back to you with the solutions.

Table of Contents