Translation Memory Services

How Translation Memory Services Reduce What You Pay Per Word

A translation memory is a database that stores every sentence we translate for you as a paired segment: the English source and its approved translation, tagged with the project, the date, and the translator who produced it. Before we start any new job, our CAT tools compare the new document against that database, segment by segment. When a sentence has been translated before, the tool retrieves it. When a sentence is close to a previous one, the tool flags the percentage of similarity and shows the linguist exactly what changed.

The commercial consequence is direct. You never pay full rate twice for the same sentence, and you pay a reduced rate for near-matches. The technical consequence matters just as much: a warning that was phrased one way in your 2023 manual stays phrased that way in your 2026 revision, because the memory serves the approved wording back to the translator instead of asking them to reinvent it.

Translation memory services, as we practice them, cover the full lifecycle: creating the memory on your first project, applying it to every subsequent job, maintaining it as your terminology evolves, cleaning it when errors creep in, and handing it over whenever you want it. Techniwords has been running TM programs for US manufacturers and engineering firms for 15 years, and the mechanics described on this page are the ones our project managers apply daily, within an ISO 17100-compliant process that always includes independent revision.

The Fuzzy Match Grid, Printed on Every Quote

Many agencies analyze your files against a TM and quietly pocket the difference. We print the analysis on the quote. Each estimate you receive from us itemizes the word counts per match band and the rate applied to each band, so you can verify the math yourself. The bands follow the standard CAT tool breakdown:

Match band	What it means	Rate applied
Repetitions	The same segment appears more than once in the new files	15% of full rate
100% matches	Segment exists in your TM with identical text and context	25% of full rate
95-99% matches	Minor difference: a number, a part reference, punctuation	40% of full rate
85-94% matches	Same sentence structure, a few words changed	60% of full rate
50-84% matches	Loose similarity, often faster to retranslate than to edit	100% of full rate
New words	No usable match in the memory	100% of full rate

Two details in that grid deserve a comment. First, 100% matches are not free: a linguist still reads each one in context, because a segment that was correct in a maintenance chapter can be wrong in a safety chapter. Second, we draw the line at 85% for discounts. Below that threshold, editing a bad match takes longer than translating from scratch, and pretending otherwise produces sloppy output. If you want to understand how these bands interact with volume pricing and minimum fees, our pricing guide walks through complete worked examples.

You Own Your Translation Memory. We Put It in Writing.

The single most common problem we solve for new clients is this: they have paid for ten years of translation, and their former agency claims the accumulated memory as proprietary work product. The client asks for their data and receives silence, a fee demand, or a degraded export with the metadata stripped out.

Our position is the opposite, and it is written into our service agreement. The translation memory built from your projects is your property. You paid for every segment in it. On request, we deliver a full TMX export, the vendor-neutral exchange format that every major CAT tool reads, with all attributes intact: creation dates, project identifiers, and status flags. There is no export fee and no notice period. We hold a working copy so your next project benefits from the memory, but the master asset is yours.

We take the same approach with termbases, which live alongside the TM and control individual terms rather than full sentences. The two assets work together, and our terminology management services page explains how the termbase side is built and validated by your own engineers.

Recovering and Cleaning a Legacy TM From a Former Vendor

When you switch providers, the memory should move with you. We handle that migration as a defined service with its own deliverables, because legacy TMs are rarely clean. A memory that has passed through several vendors and several tool versions typically arrives with duplicated segments, conflicting translations of the same source sentence, machine translation that was never reviewed, and tags broken by format conversions.

Our cleanup sequence runs in five steps:

1. Import audit. We load the TMX into a controlled environment and report what it actually contains: segment count, language pairs, date range, and the share of segments with missing or damaged metadata.
2. Deduplication. Identical source-target pairs are merged; identical sources with different targets are flagged for arbitration rather than silently overwritten.
3. Arbitration. A senior linguist resolves each conflict against your current terminology, keeping the correct variant and marking the loser as deprecated instead of deleting the history.
4. Technical repair. Broken inline tags, encoding faults, and segmentation inconsistencies are fixed so the memory produces usable matches instead of false positives.
5. Verification sample. We run a 500-segment quality sample and send you the findings before the memory goes live on production work.

Cleanup is quoted as a flat project fee based on segment count, not buried in future per-word rates. For a memory of 50,000 to 100,000 segments, the audit and cleanup typically complete in five to eight business days.

Alignment: Building a Memory From Translations You Already Have

Some clients arrive with no TM at all but with years of translated documents sitting on a server: English manuals next to their Spanish versions, datasheets in German, procedures in French. Alignment turns that archive into a working memory. Our tools pair each source document with its translation, split both into segments, and propose matches that a linguist then confirms or corrects.

The confirmation step is where alignment succeeds or fails. Automatic pairing runs at roughly 85-92% accuracy on well-structured technical documents, which sounds high until you realize that an unreviewed 10% error rate poisons every future project with wrong matches served at full confidence. Our linguists review every proposed pair on safety-relevant content and a statistical sample on the rest.

Alignment pays off fastest on document families with heavy internal repetition. Libraries of standard operating procedures are the textbook case: hundreds of documents sharing headers, step formulas, and cautionary phrasing. The same logic applies to safety data sheets, where Section 2 hazard statements and Section 4 first aid measures repeat across an entire chemical portfolio with only the substance data changing.

What Three Years of TM Management Actually Saves

The value of a memory compounds, and the curve is steep enough to plan budgets around. Here is the anonymized trajectory of a client we onboarded in 2022, a fluid handling equipment manufacturer translating manuals, procedures, and datasheets into three languages:

Year	Words submitted	Billed as new words	Effective discount
Year 1	410,000	352,000	11%
Year 2	445,000	289,000	29%
Year 3	430,000	231,000	41%

Year 1 savings come mostly from repetition inside the documents themselves. By year 3, the memory contains the client's entire product line, so every manual revision, every derivative model, and every updated procedure rides on prior work. Revision projects are the extreme case: when a manual changes 15% between versions, you pay close to 15% of the original cost, not 100%.

The savings hold only if the memory is maintained. When a client renames a product line or a regulator forces new wording, we update the affected segments and deprecate the old ones, so the discount machine does not start serving outdated text. That maintenance discipline is what separates a managed memory from a folder of TMX files.

How TM Interacts With Machine Translation and Human Revision

A frequent question from procurement teams: if machine translation is cheap, why maintain a TM at all? The two technologies do different jobs. A memory returns your own approved sentences with certainty. A machine translation engine generates plausible new sentences with no memory of what your reviewers approved last year. In our hybrid workflows, the TM always takes precedence: segments with matches above 85% come from your memory, and only the remainder is considered for machine output with human post-editing. Our post-editing workflow describes the risk triage we apply before any engine touches your content.

Whatever the source of a segment, TM match or fresh translation, the ISO 17100 workflow applies: a second, independent linguist revises the assembled target text in full. Matches are not exempt from revision, because context shifts. A segment reading "Close the valve before servicing" is a perfect 100% match whether the valve is a bleed valve or an isolation valve, and only a human reading the surrounding procedure catches when the reuse is wrong.

Where Your Memory Lives: Hosting, Security, and Access

A translation memory is a condensed copy of your documentation, sentence by sentence, so it deserves the same handling as the documents themselves. Your TM is stored on access-controlled servers, encrypted at rest and in transit, and segmented per client: your segments are never pooled into a shared memory, never used to serve matches to another customer, and never fed into any public machine translation engine or model training pipeline. Access is limited to the linguists assigned to your account, each bound by a signed confidentiality agreement before they see a single segment.

You also decide who on your side can see and receive the asset. Some clients name a single documentation manager as the export contact; others give their engineering leads read access to review how specific sentences were rendered. Under NDA, we accommodate stricter regimes as well, including project-specific memories that are archived and returned at contract close rather than retained. Whatever the arrangement, it is written into the statement of work, so the rules survive personnel changes on both sides.

Where Managed Memories Matter Most: Regulated, Repetitive Documentation

TM management earns its keep wherever documentation is both regulated and repetitive. In pharmaceutical manufacturing, batch records and cleaning procedures reuse validated phrasing that auditors expect to see rendered identically every time; a divergent translation of the same instruction across two SOPs is an audit observation waiting to happen. In aerospace programs, maintenance documentation written under ASD-STE100 Simplified Technical English is deliberately built from controlled, repeating sentence patterns, which makes it some of the highest-reuse content we process: match rates above 60% on revision cycles are routine.

Because we are an ATA and GALA member agency working from Texas with clients across the US, we also see the procurement side of this question. Companies that own their memories change vendors freely and negotiate from strength. Companies that do not are locked in by their own data. Building that independence for you is the point of the service. To see how TM management fits alongside our other engineering and language services, browse the full services catalog.

One habit we recommend to every documentation manager: schedule a TM health review once a year. Thirty minutes with our team, a fresh analysis of your latest files against the memory, and a short report on reuse rates, conflict counts, and termbase drift. It costs nothing and it tells you exactly what your linguistic assets are worth.