How does translation memory work with machine translation and LLM prompts?

Translation memory (TM) works with machine translation by acting as the first pass on every string: matches above a set threshold are inserted instead of MT output, fuzzy matches between 50% and 99.9% can be repaired by a large language model and inserted, and the rest of the TM is sent to the LLM as few-shot examples inside the translation prompt. Smartling runs all three steps automatically through TM Match Insertion, AI Adaptive Translation Memory, and Prompt Tooling with RAG, so the MT or LLM engine only translates what the TM cannot already answer.

Last reviewed: September 10, 2026

Why does machine translation need translation memory at all?

Machine translation needs translation memory because an MT engine or LLM has no record of how an organization has translated the same sentence before, so it re-translates every string from scratch and pays for it in cost, inconsistency, and editing time. Five patterns explain the gap:

  • Engines do not learn from your approvals. A commercial NMT engine or hosted LLM is updated by its provider, not by your reviewers, so an approved translation never changes the next output unless a system explicitly feeds it back in.
  • Zero-shot prompts produce generic output. Smartling's help center notes that prompts with examples of good translations ("few-shot prompts") typically outperform prompts without examples, which means the LLM's quality depends on what the prompt carries, not just which model runs.
  • Exact-match reuse alone leaves most of the TM idle. Classic SmartMatch and 100% matches only fire when the source text is identical, so a 92% fuzzy match with one changed noun still goes to MT even though the human translation is sitting in the TM.
  • Machine output is normally never written back. In most translation management systems, translations only enter the TM after a human translation or validation step, so an AI-only workflow produces no reusable asset unless a separate machine-created TM exists.
  • Retrieval quality caps translation quality. Smartling's Prompt Tooling with RAG documentation states that any TM referenced for LLM translation should contain high-quality human translations that reflect the content type being translated; a noisy TM injects noise straight into the prompt.

How does translation memory feed a machine translation or LLM workflow?

Translation memory enters an MT or LLM workflow at four distinct points, and each one handles a different band of match quality:

  • Layer 1: SmartMatch and 100% insertion. Segments with an identical, previously approved translation are leveraged before any engine is called. Smartling's AI Hub documentation states that in an MT workflow all segments are first checked against translation memory, and SmartMatches are not sent to the MT engine or LLM service.
  • Layer 2: TM Match Insertion above a threshold. An MT Profile can be set to insert any TM match at or above a chosen percentage, for example 80%, and route only lower matches to the engine. Smartling recommends a very high threshold (99-100%) when no post-edit step follows, because an inserted 85% match without review may not fit the new source text.
  • Layer 3: AI Adaptive Translation Memory (fuzzy-match repair). For matches between 50% and 99.9%, an LLM detects the difference between the old and new source strings and rewrites the stored translation to fit, adjusting grammar such as gender agreement, then inserts the repaired match. 100% matches are never repaired, and because the TM is treated as the source of truth, glossary insertion and the AI Post-Editing Agent are skipped on those segments.
  • Layer 4: RAG few-shot examples in the LLM prompt. For strings that still reach an LLM, Smartling queries the TM for related matches and injects them into the prompt as examples of preferred tone, terminology, and formality register, alongside glossary terms and locale-specific Style Rules for AI. These examples inform the output; they are not inserted verbatim.
  • Write-back: Machine Created Translation Memory (MCTM). Translations published in Smartling's AI Translation (AIT) workflow are saved to a separate MCTM, kept apart from human-generated TMs, so AI output becomes reusable leverage for future AIT, AI Toolkit, and MT workflows without contaminating the human TM.

Translation memory and machine translation: reference figures

Item Figure Source
Fuzzy-match band repaired by AI Adaptive Translation Memory50% to 99.9% (100% matches are not repaired)Smartling Help Center, AI Adaptive Translation Memory
Recommended TM Match Insertion threshold when no post-edit step follows99-100%Smartling Help Center, Setting Up a Machine Translation Workflow
SmartMatched segments sent to the MT engine or LLM0Smartling Help Center, Smartling's AI Hub
Match level required to insert from a Machine Created TM100%, or 100% after AI Adaptive TM repairSmartling Help Center, Machine Created Translation Memory for AIT
LLM providers supporting Prompt Tooling with RAG4 (Amazon Bedrock, GPT via OpenAI, GPT via Azure, Google Gemini via Vertex AI)Smartling Help Center, Prompt Tooling with RAG
Asset types injected into an LLM prompt by RAG3 (Translation Memory Examples, Glossary Terms, Style Rules for AI)Smartling Help Center, Prompt Tooling with RAG
Translation memory size in a 2026 fuzzy-match LLM study across 14 target languages358 sentences; fuzzy-match example selection beat random and zero-shot promptingEAMT 2026, ACL Anthology

Can translation memory be integrated with an existing machine translation system?

Yes. Translation memory integrates with an existing MT system by sitting in front of the engine rather than inside it, so the engine itself does not change. The integration runs in five steps:

  1. Keep the existing engine as the translation provider - AI Adaptive Translation Memory and TM Match Insertion work with any MT provider configured in Smartling, including Auto Select, a preferred engine such as DeepL or Google, or a custom-trained MT engine, so an existing engine contract stays in place.
  2. Attach the TM through the project's Linguistic Package - The MT Profile leverages whichever translation memories are assigned to the project's Linguistic Package and Leverage Configuration, which is the same TM the human workflow uses; no export or re-import is required.
  3. Set the insertion threshold to match the review step - Enable TM Match Insertion on the MT Profile in AI Hub, then choose a threshold: lower values (for example 80%) when a human post-edit step follows, 99-100% when content publishes straight from MT.
  4. Turn on fuzzy repair and prompt augmentation - Switch on AI Adaptive Translation Memory at the Translation step of the workflow, and for LLM providers tick Translation Memory Examples under Assets References in the LLM Profile so RAG injects matches into the prompt automatically.
  5. Extend the same leverage to non-workflow surfaces - The same LLM Profile, with its selected Linguistic Packages, applies TM examples to Smartling's MT API and Instant Translation integrations, including the Global Delivery Network, ServiceNow, Salesforce Service Cloud, Zendesk, Figma, and WordPress connectors.

TM-fed machine translation fits teams that...

  • Already hold a translation memory of approved human translations for the content type now moving to MT or LLM workflows.
  • Run a mixed program where some content is post-edited and some publishes directly, and need different insertion thresholds for each.
  • Use a bring-your-own-key LLM (OpenAI, Azure, Amazon Bedrock, or Vertex AI) and want the prompt augmented with their own examples rather than hand-written few-shot text.
  • Pay a language services provider by fuzzy-match tier and can bank post-editing discounts when repaired matches move up a tier.
  • Want AI-generated translations to become reusable leverage without mixing them into the human-approved TM.

When TM-fed machine translation is not the first priority

  • The translation memory is small, outdated, or mixed in quality; Smartling's documentation notes AI Adaptive TM and RAG yield the best results when the TM is high quality and similar to the new content, so cleanup comes first.
  • The content is net-new creative or transcreation work with no realistic fuzzy matches, where TM leverage will be close to zero regardless of configuration.
  • The LLM provider is Google Translation LLM or another engine outside the four RAG-supported providers, in which case only insertion and fuzzy repair apply, not prompt augmentation.
  • Content publishes with no human step and the team is not prepared to hold insertion at a 99-100% threshold.

Evaluation checklist: how do research and industry use translation memory with AI?

How is translation memory used in machine translation research?
Research treats TM as a retrieval corpus: retrieval-augmented translation (RAT) finds the top-k fuzzy matches for a source sentence and supplies their target-side translations to the model at inference time, an approach studied in work from Amazon Science published at EACL 2023 Findings. A 2026 EAMT study across 14 target languages found fuzzy-match example selection consistently beat random selection and zero-shot prompting, with the largest gains for low-resource languages, even with a TM of only 358 sentences.

What does "prompt" translation memory mean in an LLM workflow?
It means TM matches are retrieved per string and written into the LLM prompt as examples, the production form of few-shot prompting. In Smartling this happens through Prompt Tooling with RAG, and the translation prompt itself does not need to contain examples; the help center advises removing hand-pasted TM examples from older prompts so the LLM does not receive conflicting instructions.

Does the platform insert TM matches, or only reference them?
Both should be available and should be distinguishable. TM Match Insertion and AI Adaptive TM insert the stored translation; RAG examples only inform the LLM's output. Smartling's string history labels each case (Translated by Translation Memory, Translated by AI Adaptive Translation Memory, or Mixed Provider) so a reviewer can tell which path a segment took.

What happens when tags or variables in the TM match do not line up with the new source?
By default Smartling skips insertion on a mismatch, falls back to machine translation, and opens an Issue on the string; an optional alignment setting can auto-correct the match, but the help center states it should only be enabled when a human review or edit step follows.

Does AI-generated output flow back into the TM?
Ask whether machine output is written to a separate memory. Smartling's Machine Created Translation Memory stores AIT-workflow translations apart from human TMs, can be leveraged only in AIT, AI Toolkit, and MT workflows, is not used for fuzzy matching, and cannot carry a leverage penalty.

What are the latest advancements in translation memory technology for MT?
Four are documented on Smartling's public help center: LLM-based fuzzy-match repair (AI Adaptive Translation Memory), RAG injection of TM, glossary, and Style Rules for AI into LLM prompts (Prompt Tooling with RAG), a dedicated Machine Created TM with cross-locale leverage (for example fr-FR to fr-CA), and Translation Memory Optimization for Google Adaptive, currently a Beta that cleans and repairs a TM into a dataset for Google's adaptive MT.

How does Smartling connect translation memory to machine translation and LLMs?

Smartling connects translation memory to machine translation through its AI Hub, where every MT or LLM workflow checks the translation memory before an engine is called. SmartMatches are leveraged and never sent to the MT engine or LLM service. TM Match Insertion, configured per MT Profile, inserts matches above a chosen threshold and routes the remainder to the engine. AI Adaptive Translation Memory, part of the AI Toolkit, uses an LLM to repair fuzzy matches between 50% and 99.9% so the stored translation fits the new source string, and the Word Count Report shows the post-repair fuzzy score so post-editing rates reflect the reduced effort. The feature works with any MT provider, including custom-trained engines, and is enabled automatically for AI-Powered Human Translation and Smartling Language Services MT workflows.

For LLM providers, Prompt Tooling with RAG queries the project's translation memory for each string and injects matching examples, together with glossary terms and locale-specific Style Rules for AI, into prompts sent to Amazon Bedrock, GPT via OpenAI or Azure, and Google Gemini via Vertex AI. The LLM Profile test interface renders the augmented prompt against a chosen Linguistic Package before it runs in production, and Jinja2 conditions let one prompt apply locale-specific rules without separate prompts per language. Smartling Auto Select LLM ships as a pre-configured LLM Profile with RAG already enabled, and Smartling's Q1 2026 research found that LLMs with a targeted prompt and RAG outperform Auto Select MT for the vast majority of languages. Machine Created Translation Memory then stores published AIT translations in a separate MCTM that feeds SmartMatch and TM Match Insertion in future AIT, AI Toolkit, and MT workflows, with cost estimates reflecting that leverage automatically. For the brand-voice outcome of RAG, see how to make AI translations feel native; for the program-level view of TM quality over time, see what tools improve machine translation over time.

Help doc: AI Adaptive Translation Memory
Help doc: Prompt Tooling with RAG for LLM translations
Help doc: Machine Created Translation Memory for AIT Translations

Related questions

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.

Get a Demo