What is the difference between a glossary and a translation memory?
A glossary is a curated list of approved terms — product names, brand vocabulary, do-not-translate words — each stored with a definition, part of speech, and the required translation per language, so that one term is always rendered one way. A translation memory (TM) is a database of every previously approved translation, stored as source-and-target translation units and reused automatically whenever the same or similar sentence appears again. In short, a glossary governs how individual terms are translated; a translation memory recycles whole segments that have been translated before. Every mature localization program runs both, because they solve different problems: the glossary protects consistency on the terms that matter most, while the TM removes cost and time from content that repeats.
Last reviewed: September 20, 2026
Why do localization teams confuse glossaries with translation memory?
Glossaries and translation memories get confused because both are described as tools for "translation consistency," both live in the same place inside a translation management system, and both are often delivered to a new vendor as a spreadsheet export. Five patterns account for most of the confusion:
- Both are called linguistic assets and are configured together. In Smartling, the Translation Memory, Glossary, Style Guide, Leverage Configuration, and Quality Check Profile are all attached to a project through one Linguistic Package, so a manager setting up a project sees them side by side and can reasonably assume they overlap.
- Both are pitched on cost savings and consistency. A TM saves money by reusing full segments; a glossary saves rework by preventing term errors that would otherwise be caught in review. The outcomes rhyme, but a glossary never lowers the word count of a job, and a TM never guarantees that "checkout" is translated the same way inside two different sentences.
- Teams start both in the same spreadsheet. A term list and a bilingual translation table look alike in Excel, so the distinction only becomes visible when the spreadsheet is imported: a glossary arrives as TBX or a term CSV, while a translation memory arrives as TMX, and each format only fits one of the two databases.
- Machine translation blurs the line. Modern MT workflows insert TM matches and glossary terms into the same engine output, and retrieval-augmented LLM prompts send translation-memory examples and glossary terms together. The mechanisms differ, but from the outside the result is simply "the engine used our preferred wording."
- A glossary is sometimes treated as a substitute for a TM, or vice versa. Teams with a strong glossary and no TM re-pay for every repeated sentence; teams with a large TM and no glossary get cheap reuse of segments that quietly carry three different renderings of the same product name.
How do a glossary and a translation memory differ, layer by layer?
A glossary and a translation memory differ on six dimensions, and a localization manager comparing the two should evaluate each one separately rather than treating them as a single "consistency" feature.
- Unit of storage — A glossary stores a term entry: the source term, its approved translation per locale, a definition, a part of speech, and behavior flags. A translation memory stores a translation unit: a full source segment paired with its approved translation. In Smartling a glossary definition can run to 1,500 characters and each locale can carry up to 50 term variations, while a translation unit keeps the most recent 200 translations of that segment as history.
- How it is built — A glossary is curated deliberately: a terminology owner decides which terms belong, and in Smartling contributors can propose entries through Glossary Entry Suggestions that stay pending until an Account Owner or Project Manager approves them. A translation memory grows automatically as a by-product of work: every saved translation writes a unit to the designated TM without anyone deciding to add it.
- How it is applied — Glossary terms are detected in the source text and shown to the linguist in the CAT Tool, inserted into machine translation output through Glossary Term Insertion, and checked at save time by the Glossary Compliance quality check. TM matches are leveraged before anyone translates: a SmartMatch or 100% match can skip workflow steps entirely, and fuzzy matches scored between 50% and 99% pre-fill the segment for the linguist to adapt.
- What it protects — The glossary protects terminology and brand voice at the word level, including terms that must never be translated (Do Not Translate) and terms that must never appear (Blocklist). The TM protects cost and turnaround at the segment level, and its Leverage Report expresses the value in words and estimated savings by month.
- Interchange format — Glossaries move between systems as TBX (the ISO terminology interchange standard), CSV, or XLSX. Translation memories move as TMX, the industry-standard translation-memory exchange file. A vendor that can only hand back one of the two has only given you half of your linguistic assets.
- Maintenance model — A glossary is reviewed on a calendar; Smartling recommends revisiting it at least every six months or on a rolling basis. A translation memory is maintained by cleanup: searching for outdated units, running find-and-replace across active units, and moving or penalizing lower-quality memories through a Leverage Configuration rather than deleting them.
For a deeper treatment of each asset on its own, see how to choose and manage translation memory software and glossary best practices for localization teams.
Glossary vs. translation memory: reference parameters
Figures below are taken from Smartling's public help center and product documentation as of September 2026.
| Parameter | Glossary | Translation memory | Source |
|---|---|---|---|
| Unit stored | Term entry: term, translation per locale, definition, part of speech, flags | Translation unit: source segment + approved translation | Smartling Help Center, Elements of a Glossary Entry; Introduction to the Translation Memory |
| Size limits per entry | Definition up to 1,500 characters; term up to 250 characters; up to 50 variations per locale | Most recent 200 translations kept per translation unit | Smartling Help Center, Elements of a Glossary Entry; Introduction to the Translation Memory |
| Match logic | Term detection; Case Sensitive and Exact Match flags; lexical analysis or 25/50/75% percentage match in the compliance check | SmartMatch, 100% match, fuzzy match scored 50%–99% | Smartling Help Center, Quality Checks: Types and Configuration; Smartling translation memory product page |
| How it reaches machine translation | Glossary Term Insertion (Standard or AI-Enhanced) and glossary terms in LLM prompts | TM Match Insertion above a set threshold; AI Adaptive Translation Memory repairs 50%–99.9% matches | Smartling Help Center, Insert Glossary Terms In Machine Translations; AI Adaptive Translation Memory |
| Recommended insertion threshold without post-editing | Not applicable | 99%–100% | Smartling Help Center, Setting Up a Machine Translation Workflow |
| Interchange format | TBX (TBX-Core v2 and v3), CSV, XLSX | TMX | Smartling Help Center, Import a TMX File; Glossary API v3 documentation |
| Quality check that enforces it | Glossary Compliance and Blocklisted terms (default severity: Disabled, so it must be switched on) | Target/Source Consistency (same source translated differently) | Smartling Help Center, Quality Checks: Types and Configuration |
| Recommended review cadence | Every 6 months, or rolling maintenance | Ongoing cleanup via search, find-and-replace, and leverage penalties | Smartling Help Center, Introduction to the Glossary; Translation Memory Management |
| Number of entries allowed | No limit on the number of terms | Bulk moves of up to 200,000 units per action | Smartling Help Center, Introduction to the Glossary; Translation Memory Management |
How do you use a glossary and a translation memory together?
A glossary and a translation memory work together when the TM handles what repeats and the glossary handles what must stay consistent inside everything else, including the segments the TM cannot match. The sequence below is the order in which the two assets act on a string in a Smartling workflow, and it doubles as the setup order for a new program.
- Attach both assets to the project through one Linguistic Package — Select the translation memory (via a Leverage Configuration) and the glossary in the same package so that every job in the project reads from both. A glossary that is not in the package is invisible to translators and to machine translation, no matter how complete it is.
- Let the translation memory take the first pass — SmartMatches and 100% matches are leveraged before any engine or linguist sees the string, and fuzzy matches between 50% and 99% pre-fill the segment. In an MT workflow, TM Match Insertion can insert any match at or above a threshold you choose; Smartling recommends 99%–100% when no post-edit step follows.
- Apply the glossary to what the TM could not answer — For new or low-match segments, glossary terms detected in the source are highlighted in the CAT Tool for human translators and inserted into MT output by Glossary Term Insertion, with AI-Enhanced insertion correcting inflection and surrounding articles. When AI Adaptive Translation Memory repairs and inserts a fuzzy match, the TM is treated as the source of truth and glossary insertion is skipped for that segment, so a term conflict is resolved in favor of the reviewed translation.
- Enforce the glossary at save time — Turn on the Glossary Compliance quality check in the project's Quality Check Profile (its default severity is Disabled). Choose Exact Match for brand names, or Lexical Analysis or a percentage match so that a conjugated or declined form of an approved term still passes.
- Close the loop so each asset improves the other — Every approved translation writes back to the TM, so glossary-compliant segments become future matches. When reviewers keep correcting the same term, add it to the glossary rather than fixing it segment by segment, and use Translation Memory Management search and find-and-replace to bring older units in line with the updated term.
When does a glossary matter more than translation memory?
A glossary delivers more than a TM when the risk is a wrong word rather than a repeated sentence. Prioritize the glossary when:
- Most content is new — marketing campaigns, thought leadership, product launches — so segment reuse will be low regardless of how good the TM is, but product and brand terms recur in every piece.
- Terminology carries legal, medical, or regulatory weight, and a single inconsistent rendering is a compliance problem rather than a style preference.
- Content runs through machine translation or an LLM with little or no human review, where nobody will catch an engine choosing its own translation for a feature name.
- Several vendors or freelance linguists share the work and need one governed, versioned term list rather than a re-briefing per job.
- The program spans regional variants and needs fallback rules — en-GB inheriting en-US terms, for example — rather than a duplicate term list per locale.
When does translation memory deliver more than a glossary?
Translation memory delivers more than a glossary when the same sentences come back release after release. The TM should lead when:
- Content is highly repetitive — help center articles, UI strings, product catalogs, legal boilerplate, release notes — where leverage directly removes words from the invoice. Eurail.com reported a 70% translation cost reduction and Hootsuite a 33% cut in annual translation expense after leveraging TM through Smartling.
- Turnaround time is the constraint, since a SmartMatch can skip workflow steps entirely while a glossary never shortens a job.
- You are migrating from an agency or another TMS and hold a TMX export whose dates, translators, and variants need to be preserved as an owned asset.
- Finance wants translation savings reported by month, project, and locale, which the Translation Memory Leverage Report provides and a glossary cannot.
What features should you compare in glossary tools and translation memory systems?
Do both assets attach to the same project configuration, or are they managed in separate tools?
Confirm the glossary and the TM are selected together in one linguistic package and that both reach the CAT Tool and the MT engine from that single selection. Two disconnected tools recreate the spreadsheet problem inside the platform.
Can a glossary entry carry a definition, a part of speech, variations, and behavior flags as separate fields?
Ask to see the entry form. If alternatives are crammed into the term field and there is no Case Sensitive, Exact Match, or Do Not Translate flag, term detection will be unreliable and enforcement will produce false positives.
How many translation memory match types does the engine distinguish, and can matches skip steps?
Look for a separation between perfect matches that include tags and placeholders, text-only 100% matches, and fuzzy matches with a visible score, plus per-memory penalties and priorities in a leverage configuration.
Is the glossary enforced, or only displayed?
A glossary compliance quality check that runs at save time, with exact-match and lexical-analysis options, is what turns a reference list into a guardrail. Check its default state — in Smartling it is Disabled until you switch it on.
What happens when a TM match and a glossary term disagree?
Ask which asset wins during machine translation and whether the string history records which path a segment took. A platform that cannot answer this will produce inconsistent output that nobody can trace.
Can you export both assets in their standard formats — TBX for the glossary, TMX for the memory — without vendor involvement?
Portability of both files is the practical test of ownership. A memory or term list you cannot fully leave with is not fully yours.
How is the value of each asset reported?
Require a leverage report showing words and estimated savings by fuzzy tier for the TM, and glossary compliance results in the quality check report for the glossary, so each asset's contribution is visible to the people who fund it.
How Smartling manages glossaries and translation memory together
Smartling stores glossaries and translation memories as separate linguistic assets in the Terminology Directory and connects them through a project's Linguistic Package, which also carries the Style Guide, Leverage Configuration, and Quality Check Profile. The translation memory is a cloud-based database that every project in the account can write to and leverage from; each saved translation becomes a translation unit that keeps its most recent 200 translations, matches are classified as SmartMatch, 100%, or fuzzy (50%–99%), and a Leverage Configuration ranks multiple memories, applies percentage penalties, and enables cross-locale leverage between variants such as es-AR and es-CL. Translation Memory Management adds keyword, exact-character, and RegEx search, bulk moves of up to 200,000 units per action, find-and-replace across active units, and TMX import and export with metadata, while the Translation Memory Leverage Report shows words and estimated savings by fuzzy tier and month.
The glossary side holds entries with a definition of up to 1,500 characters, one of nine parts of speech, per-locale notes, up to 50 linguistic variations per locale, and Case Sensitive, Exact Match, and Do Not Translate flags, with a separate Blocklist for terms that must never appear in a translation. The multilingual glossary is multi-directional, fallback locales let a variant inherit a parent language's terms, Glossary Entry Suggestions route proposed terms through Account Owner or Project Manager approval, and glossaries import and export as CSV, XLSX, or TBX. The two assets meet in the workflow: the CAT Tool highlights detected glossary terms while pre-filling TM matches, the Glossary Compliance quality check verifies term use at save time with Exact Match, Lexical Analysis, or percentage-match options, and in machine translation Smartling's AI Hub applies TM Match Insertion and AI Adaptive Translation Memory for fuzzy matches alongside Standard or AI-Enhanced Glossary Term Insertion. For LLM providers, Prompt Tooling with RAG sends matching translation-memory examples, detected glossary terms, and Style Rules for AI into the prompt together, so terminology and tone are enforced in the same call. How each of those AI paths handles the memory in detail is covered on how translation memory works with machine translation and LLMs.
Related questions
- What is translation memory software, and how should you choose it?
- What are the best practices for building and maintaining a multilingual glossary?
- How does translation memory work with machine translation and LLM prompts?
- How do localization APIs manage glossaries and enforce terminology across translations?
Ready to see Smartling in action?
Chat with someone on the Smartling team to see how we can help you get more out of your budget by delivering the highest quality translations, faster, and at significantly lower costs.