Translation memory: How does it work and how to make the most of it

Q: How should teams manage and maintain a translation memory over time?

Six practices keep a TM reliable as a project scales. Use project-specific TMs for different content types — separate databases for UI copy, marketing content, legal documentation, and game narrative prevent mismatched suggestions from bleeding across categories. Create a master TM for only the best approved translations you are confident in reusing. Synchronize locales consistently across platforms to ensure the TM matches correctly across all surfaces. Schedule regular clean-up cycles to remove outdated, incorrect, or duplicate entries. Assign an admin responsible for monitoring correct usage and resolving inconsistencies. Allocate onboarding time for new translators and reviewers — TM settings and workflows have a learning curve that affects output quality if rushed.

Q: What features should teams prioritize when evaluating translation memory software?

Eight capabilities matter most: file format support for import and export (TMX, CSV, XLSX) for portability and bootstrapping; a TM editor for direct review and correction of stored entries; visible match percentage data to help translators decide whether to apply or rework a suggestion; metadata support for tagging content by type, platform, or style; alternative translation support for context-dependent variants; concordance and wildcard search for fast term lookup; role-based access controls for translators, reviewers, and admins; and configurable match-handling settings that define which tiers trigger automatic population versus human review, tailored to each content type's risk profile.

What’s covered:

What is a translation memory?

A translation memory is a database that stores segments of translated content. It can be limited to pairs of content—in the original language and one target language—or include a range of target languages. Entries in a translation memory are logged with their source language — the original text — as well as the translated versions in various target languages.

Under a content segment, there can be a word, a phrase, a sentence, or a whole paragraph. However, paragraphs are quite rarely used, as it’s not common to have a large piece of identical content reused.

How is translation memory different from a termbase or glossary?

To clarify terminology:

A glossary is a list of terms used in a product that features their explanations and approved translations.
A termbase is also a list of terms but it might feature different variations of terms and their translations, as well as the context they can be used in.

While glossary and termbase tools help ensure consistency just like translation memory, the latter is a more complex mechanism that is meant not just as a reference but as a way of efficiently reusing existing translations.

What are the benefits of a translation memory?

Translation memory makes your localization workflow faster, better, and less expensive. Once an entry is logged in your translation memory, you’ll never need to translate it again. That’s why any localization software worth using will come with this feature built in.

The benefits of using a translation memory include:

Reduced costs: Translation memories eliminate redundant work by preventing translators from having to retranslate earlier material. They’ll be able to apply their billable time solely towards translating new text.
Increased accuracy: The translations stored in your memory will have already been verified the first time the translation was performed.
Greater consistency: Translating terms and phrases the same way makes your content easier to understand while boosting your brand — both essential for engaging with your audience.
Faster turnaround: By focusing just on untranslated text, your translation team will greatly improve their efficiency. This can massively streamline your localization workflow and shorten your project timelines.

How does a translation memory system work?

The main purpose of a translation memory is to automatically reuse approved translations. It will still require some manual actions to check if the match is correct for each new use and context. Here’s how it works:

Every time translators input new content into a localization management system, the system detects if some segments are identical or similar to what is already approved and stored.
When the translation memory tool finds a match, it calculates the percentage of similarity to attribute the match to a particular type (more on that later). In the settings, you can specify how you’d like to process different types of matches. For instance, you might choose to ignore formatting, and the system will consider two segments identical even if they differ in the usage of capital letters or bold font style.
When the match is identified, a translation memory either automatically fills out the string (which you can review and approve) or provides you with auto-generated suggestions once you start filling the string in each target language.

In Gridly, translation memory can automatically populate all target language columns. They will be visually marked by “TM” so that you can further check the relevance of identical translations, filter them out, and approve them.

>> Learn more about how Gridly can help you with your localization workflow.

Additionally, when you start editing text in a target language column, Gridly automatically suggests reusing existing translations. Once you click on it, the system will show you auto-generated suggestions, and you can choose and approve the best-fitting one.

Auto-suggestions by the translation memory in Gridly

Alternative translations

You might need several variations of translations for the same phrase. For instance, you might need phrases of different lengths for different UI elements or simply want to have synonyms in different contexts.

A translation memory in Gridly supports alternative translations. You can add multiple variations for the same source text, and each will be assigned a different ID for reference. When working on localization, your translators will view all available options and choose the best one that suits each particular context.

Alternative translations in Gridly’s translation memory

Search within a translation memory

Sometimes, you might want to recheck something in the translation memory database. To do so, you’ll need your localization tool to support flexible search and filtering options:

Concordance search. When you need to look for all instances of a particular word or phrase, a concordance search will highlight the results. The search can be performed in the source language as well as in any target language.
Wildcard search. For greater flexibility, you can use different wildcard settings when you need to search for different variations of a word or parts of compound words.

Different types of matches in a translation memory

Different software systems have their differences in what match tiers they use and how they process them, but generally, there are perfect matches, 100% matches, and fuzzy matches.

Perfect matches, also called 101% or context matches, mean that a piece of content is both textually and contextually the same. This is the type of translation that you want to lock in and reuse automatically, without manual review. However, make sure that the translations you lock in are actually the best possible options and are approved by the people in charge.
100% matches mean that the segment completely matches the existing text but the context is different. For example, the identical phrase is surrounded by segments that are not identical to previous translations. In most cases, the 100% matches will be reused as suggested by a translation memory, but it still makes sense to double-check the context manually. However, it depends on the nature of the product you’re translating. If you know that there can hardly be any ambiguities with certain terms, you can set the translation memory to automatically replace 100% matches.
Fuzzy matches are distinguished when the similarity is under 100%. The calculation depends on a range of factors, including word order, formatting, string length, punctuation marks, etc. Usually, fuzzy match classification starts at the 70-75% match as anything lower probably requires being translated from scratch. For example, in a phrase or sentence of around 10 words, you’ll get an 85-95% match if there’s a difference in one word and a 75-85% match if two words are different.

Different types of matches in a translation memory

Key use cases for translation memory

Translation memories are best used when consistently translating between specific source and target languages, and with a large amount of repeated strings. A translation memory can be especially helpful with the following types of content:

Legal and financial documents: Official paperwork often contains lengthy segments of text repeated across documents. And the serious nature of this content makes it critical to maintain accuracy. Translation memories can help on both fronts.
Customer-facing content: From industry-specific language to branded terms, a business will often have plenty of content set in stone. Use a translation memory to ensure consistency across an entire brand experience, including your website, any third-party collateral, and app.
App localization: To prevent users from getting confused, UX writers name buttons, menu items, and other software elements with exacting specificity. Use a translation memory to repeat these terms throughout an app or other software.
Game localization: Video games, especially narrative games like RPGs, can include hundreds of characters and locations along with thousands of item names and descriptions. Use a translation memory to keep everything consistent for your players.

How to create a translation memory?

The process of creating a translation memory will depend on the localization tool you’re using. Here’s how you can do it in Gridly:

On the Translation Memory page within the tool, select “Create TM” and name it.
Assign your translation memory to a particular localization project. Note that the languages are defined when creating a Grid (which is a spreadsheet-like interface for real-time translation). You can add several target languages.
If you already have approved translations that you want to reuse, upload them to Gridly in the .tmx, .csv, or .xlsx format. This way, your existing translations will be stored in the translation memory. To apply those translations to your existing strings, sync the TM—and it will fill out the translations based on the uploaded data.
Everything that is translated in a Grid will automatically go to the translation memory system and be processed for matches, except translations created by machine translations and AI-powered tools in case they weren’t edited by a translator.
You can edit your translation memory, remove or add terms, search for translations, add multiple alternative translations, etc.

Translation Memory editing in Gridly

How to manage a translation memory

The more content you have in your localization project, the harder it might be to search for something when needed. Plus, there might be inconsistencies or errors if your product is available on multiple platforms, and locales for those different platforms aren’t unified in a localization tool.

To maintain a clear database, follow these translation memory management rules:

Use the settings that work for your project. Based on the nature of the product you’re localizing, specifics of the design, or other nuances, it might make sense to adjust the settings of the translation memory software. For example, make certain words and phrases ( unchangeable brand names, slogans, or standard UI copy) always automatically translated based on stored translations. Or, set exact matches with different font style tags to always be manually reviewed if font styling might completely change the context.
Create different translation memories within one project. It might make sense to create separate translation memories for the product’s UI, marketing content, documentation, etc. You can also create separate databases if the product you’re localizing has different versions—for instance, a mobile app and a website, or a PC and PS game.
Create a master translation memory to put only approved translations there. Use different project-based translation memories and move only the best translations that you’ll definitely reuse to the master translation memory.
Synchronize locales between projects. Platform-specific versions of a product might be stored inconsistently in a translation management system, leading to problems with translation memory, so make sure that the tool you’re using syncs between different platforms and correctly assigns each target language.

In Gridly, you can easily avoid problems with syncing the content between different app platforms (iOS, Android) or game consoles (Xbox, PS). You can add various types of metadata and content to assign a platform for each source text. You can also work with alternative translations in the translation memory if you want to have text variations based on the platform.

Set up a clean-up and update schedule. Overcrowded translation memories will lead to more errors and inconsistencies, so you should regularly maintain your TM databases. A clean-up can include removing incorrect or outdated translations and merging duplicates. Exclude all segments whose quality you’re not sure about, as they will negatively affect the quality of automated translations. If the system allows you to filter fields by the date they were last modified or used, you can easily spot old content and remove it if necessary. Also, regularly check if there are technical issues like escape characters that can cause problems for translators.

Get more translation memory maintenance tips

As for translation memory updates, you might want to regularly do so for machine-translated segments to be included. For instance, Gridly doesn’t store automated translations in the translation memory unless you manually update the database.

Have an admin monitoring correct usage. Assign an admin role for the person who’ll be overseeing the management of the translation memory. An admin should regularly check if correct language pairs are used and if all files uploaded into the system match a given context.
Provide tutorials and assistance to translators and reviewers. A translation memory system has a certain learning curve, especially for localization specialists who haven’t used a similar tool or are moving from one tool to another. Give your localization team time and resources to familiarize themselves with the software and adjust the settings according to the project. Last but not least, choose a tool that is intuitive in use and highly customizable—this way, people will spend less time learning how to use it and dedicate more effort to the actual translation.

What to look for in translation memory software

A translation memory can be a blessing, but only if it fits your needs. Here are the major things to pay attention to when choosing a localization system with a translation memory:

Import and export with popular file types support. It’s essential to be able to sync existing records with your translation memory, as well as upload files to individual strings to clarify context. Search for the localization management systems that support .xtm, .scv, .xlsx, etc.
TM editor. With a flexible translation memory editor, you can easily review and edit translations in each target language.
Data on the match percentage. It can be helpful to see the match type or percentage to quickly understand if certain strings need to be manually reworked. It’s also crucial to have flexible settings for different types of translation memory matches (in some instances, you might need 100% matches to be automatically filled out, while in others, you might want those 100% matches to be left for human review).
Metadata. To organize the content in translation memory software, you can use various metadata attributes. For example, metadata helps assign types of content (UI copy, documentation, etc.), platforms where the product is released (iOS, Android, etc.), style (formal, informal, slang, etc.), or anything else that makes sense for a particular product.
Alternative translation support. Since the same word or phrase can be used multiple times in different settings, it’s helpful to have a number of options to choose from for each particular context. Search for a tool that supports adding and editing multiple alternative translations in target languages.
Flexible concordance search. Concordance search will help you quickly find matches for phrases, words, and parts of words when you need to verify certain strings or check grammar.
Role-based access rights. Since different specialists will be working on localization—translators, proofreaders, managers, and designers—you need to assign different roles in your translation memory system. It’s particularly important to have an admin role in place: the admin will oversee the processes, monitor translation memory clean-ups, and prioritize different translation memories created for a project.
Customizable settings. Translation memory software should be tailored to your needs. For example, the text fields and metadata types for an ecommerce app will be very different from the text fields and metadata types for a game. The same goes for the settings where you define how the system should treat different types of matches, formatting issues, and other nuances.

Essential features when choosing a localization system with a translation memory

How does translation memory make AI translation more accurate?

Translation memory improves AI translation quality because it gives the model the context it needs to make consistent decisions. Without it, an AI model translating in isolation has no way of knowing that your game’s protagonist is always referred to by a specific name variant, that a UI button label has an approved short form for character-limited strings, or that a particular phrase carries a tone that differs from its literal meaning. The model guesses. TM removes the guesswork.

When TM data is passed to the model alongside the source string, the output aligns with decisions your team has already made and approved. Terminology stays consistent across updates. Brand voice carries through new content. Returning players see the same language they recognise from previous versions.

This is the most direct productivity gain from connecting TM to an AI translation workflow: approved translations act as few-shot examples, showing the model exactly how your team handles specific phrases, without requiring additional prompt engineering for each one.

What are the challenges of integrating translation memory into AI prompts?

The concept is straightforward, but the implementation is not. Most localization teams find that connecting TM data to AI translation requires more manual work than expected - and that the quality of the output depends almost entirely on how well that connection is built.

Common challenges include:

Prompt preparation overhead. To pass TM matches to a model, someone has to extract the relevant segments, format them correctly for the prompt, and decide how many matches to include before the context window becomes too large. Done manually, this adds significant setup time per job.
Inconsistent context across strings. A batch of strings may have a mix of 100% matches, fuzzy matches, and no matches at all. Treating all of them the same way in a prompt produces uneven output - over-relying on weak TM matches in some cases, ignoring strong ones in others.
No glossary or length enforcement. Even when TM data reaches the model, there is usually no mechanism to enforce glossary terms or character limits unless those constraints are explicitly written into the prompt. This is easy to miss and difficult to audit across large batches.
Output mapping. Once the model returns a translation, getting that output back into the right column of the right string record - reliably, at scale - typically requires custom scripting or manual copying.
Prompt drift. When prompts are edited between runs, earlier jobs become difficult to reproduce. Without version control on the prompt itself, debugging inconsistencies across batches becomes guesswork.

Each of these challenges is manageable in isolation. Together, they mean that teams either invest significant engineering time building a reliable pipeline, or they run AI translation as a rough first pass and absorb the post-editing cost downstream.

Frequently asked questions

What is translation memory?

Translation memory is a database that stores segments of previously translated content — words, phrases, sentences, or short paragraphs — paired with their approved translations in one or more target languages. Every time a translator completes and approves a new translation, it is logged in the database. When that same or similar content appears again in any future project, the system surfaces the existing translation automatically rather than requiring the work to be done from scratch. Translation memory is one of the few localization tools where value compounds over time: the larger and better-maintained the database, the greater the efficiency and consistency gains on every subsequent project.

How is translation memory different from a glossary or termbase?

A glossary is a reference list of terms used in a product, along with their explanations and approved translations — it tells translators what specific words mean and how they should be rendered. A termbase is similar but may include multiple variants of a term, different contextual uses, and additional metadata about each entry. Translation memory is a more complex mechanism: rather than providing a reference for individual terms, it stores and automatically reuses entire approved segments — full phrases, sentences, or UI strings — across projects. All three tools work together to enforce consistency, but translation memory operates at the segment level and is the primary driver of translation speed and cost reduction in ongoing localization programs.

What are the different types of matches in a translation memory and how are they handled?

There are three main match tiers. Perfect matches — also called 101% or context matches — occur when a segment is textually identical to a stored entry and appears in the same surrounding context. These can typically be applied and locked automatically without manual review. Exact or 100% matches occur when the segment text is identical to a stored entry but the surrounding context differs. These are usually applied automatically as well, though manual review is advisable when context could affect meaning. Fuzzy matches occur when similarity falls below 100% — typically classified starting at 70–75%, with anything lower generally treated as new content for full translation. A ten-word sentence where one word changes might score 85–95%; two changes might score 75–85%. Match percentage settings can be configured per project to control which tiers trigger automatic population and which require human review.

What is the difference between a 100% match and a perfect match?

A 100% match means the segment text is word-for-word identical to a stored entry, but the surrounding strings are different. A perfect match — sometimes called a 101% or context match — means the segment is both textually identical and contextually identical: the strings immediately before and after it also match what was previously translated. The distinction matters because context can affect meaning, particularly for short UI strings or pronouns that change based on what precedes them. Perfect matches carry higher confidence that the stored translation will be correct without any review; 100% matches usually will be correct but are worth a quick human check in ambiguous cases.

What types of content benefit most from translation memory?

Content with high repetition rates and stable terminology benefits most. Legal and financial documents contain lengthy standard clauses that appear verbatim across multiple files — translation memory prevents these from being retranslated every time and eliminates the risk of inconsistency between documents where consistency is legally important. Customer-facing product content — websites, apps, in-product copy — often reuses the same branded terms, navigation labels, and standard UI strings across multiple surfaces. App and software localization benefits particularly because button labels, menu items, and system messages are named with precision and must remain identical wherever they appear. Game localization — especially in narrative RPGs with large item databases and recurring terminology — uses translation memory to maintain consistency across thousands of strings and multiple content updates.

What are alternative translations and when should teams use them?

Alternative translations allow multiple approved variants of the same source segment to be stored in the translation memory simultaneously, each assigned a unique ID. This is useful when the same phrase needs to render differently depending on context — for example, when a UI label must fit a character limit on mobile but can be longer on desktop, or when the same term is used both formally in documentation and informally in in-game dialogue. When a translator encounters a segment with multiple stored alternatives, the system presents all available options so they can select the most appropriate one for the current context. Without alternative translation support, teams must either compromise on one version for all contexts or manage variations manually outside the TM system.

How do you create and set up a translation memory for a localization project?

The setup process typically involves creating a named TM database within the localization platform, assigning it to a project and specifying the source and target language pairs, and importing any existing approved translations in a supported format — TMX, CSV, or XLSX are the most common. Once imported, syncing the TM against existing strings in the project populates those strings with matches from the uploaded data. From that point, every new translation approved by a human translator is automatically added to the database, building its value incrementally. Machine translation output is typically not ingested automatically — those segments enter the TM only after a human translator has reviewed and approved them, preventing unverified AI output from contaminating the database.

How should teams manage and maintain a translation memory over time?

Six practices keep a TM reliable as a project scales. First, use project-specific TMs for different content types — separate databases for UI copy, marketing content, legal documentation, and game narrative prevent mismatched suggestions from bleeding across content categories. Second, create a master TM for only the best approved translations that you are confident in reusing, separate from project-level working databases. Third, synchronize locales consistently across platforms: if a product exists on iOS, Android, and web, ensure all platform-specific translations use consistent language codes so the TM matches correctly across all surfaces. Fourth, schedule regular clean-up cycles to remove outdated, incorrect, or duplicate entries — overcrowded TMs produce more noise than signal. Fifth, assign an admin responsible for monitoring correct usage, auditing language pairs, and resolving inconsistencies. Sixth, allocate onboarding time for new translators and reviewers — TM settings and workflows have a learning curve that affects output quality if rushed.

What features should teams prioritize when evaluating translation memory software?

Eight capabilities matter most. File format support for import and export — TMX, CSV, XLSX — is essential for portability between tools and for bootstrapping a new TM from existing approved translations. A TM editor that allows direct review and editing of stored entries prevents errors from compounding. Visible match percentage data for each suggestion helps translators make faster decisions about whether to apply or rework a match. Metadata support allows content to be tagged by type, platform, style, or any project-specific attribute that aids organization and filtering. Alternative translation support handles the reality that the same source phrase often needs different target variants in different contexts. Concordance search — including wildcard search — enables fast lookup of any term across the entire database. Role-based access controls ensure translators, reviewers, and admins have appropriately scoped permissions. Configurable match-handling settings allow teams to define which match tiers trigger automatic population and which require human review, tailored to the risk profile of each content type.

How does translation memory improve AI translation accuracy?

Translation memory improves AI translation quality because it gives the model explicit context about decisions your team has already made and approved. Without TM, an AI model translating in isolation guesses at terminology, tone, and phrasing — and those guesses are inconsistent across batches and across time. When TM matches are passed to the model alongside the source string, approved translations act as few-shot examples that show the model exactly how your team handles specific phrases. Terminology stays consistent across content updates. Brand voice carries through new strings. Returning players or users see language that is consistent with previous versions. The most direct benefit is eliminating the gap between AI output and house style — TM data does more to close that gap than prompt engineering alone.

What are the main challenges of integrating translation memory into AI translation workflows?

Five challenges consistently slow teams down. Prompt preparation overhead is significant: to pass TM matches to an AI model, someone must extract relevant segments, format them correctly for the prompt, and decide how many matches to include before the context window fills. Inconsistent context across string batches — a mix of perfect matches, fuzzy matches, and no matches — produces uneven output if all strings are treated identically in the prompt. Glossary and character limit enforcement are absent unless explicitly built into the prompt, which is easy to miss and hard to audit at scale. Output mapping — getting the model’s translations back into the correct columns of the correct strings without manual copying — typically requires custom scripting. Prompt drift between job runs makes debugging inconsistencies across batches difficult without version control on the prompt itself. Each challenge is manageable individually; together they represent significant engineering investment or ongoing post-editing cost.

How does Gridly’s translation memory work?

Gridly’s translation memory is integrated directly into the content management workflow rather than operating as a separate tool. When content is translated in a Grid, the system automatically checks stored entries and visually marks cells populated via TM so translators can review and approve them. When a translator begins editing a target language cell, the system surfaces suggestions from the TM automatically — including alternative translations — so the translator can select the best-fitting option for each context. The TM can be searched using concordance and wildcard search in both source and target languages. Import and export are supported in TMX, CSV, and XLSX formats. Machine translation output is not ingested automatically — it enters the TM only after human review and approval, protecting the database from unverified AI content. Multiple TMs can be maintained within a single project for different content categories, and an admin role can be assigned to oversee clean-up, language pair consistency, and overall database quality.

Localize faster with a translation memory

Digital products tend to have a lot of repeated content, and if you’re manually translating identical or similar strings, it might hurt the consistency of the brand voice, the correctness of terminology, and the overall quality of localization. Besides, it might simply be boring for translators. After all, when you get each certain translation right, you want to make sure that the best translation is reused.

With a flexible translation memory tool, you might have matching phrases automatically placed in each target language. Plus, you’ll have the possibility to store multiple alternative translations and choose the most appropriate one in each given context.

With Gridly, you can work with a translation memory in all target languages at once and adjust the tool to your preferences. Learn more about how to use it to continuously translate your strings.