Translation Memory, rewired for AI.
Legacy TM works by character similarity — it finds strings that look like your source text and substitutes them. The approach here is different: past translations are retrieved by meaning, not by character overlap, and injected as style references rather than substituted word-for-word.
Translation Memory in Flixu retrieves past approved translations by semantic similarity — meaning, not character overlap. Matches are injected into the translation request as reference material, not blindly substituted. A fuzzy match found in one section of a document improves translation quality throughout the document, not just at the point where the match appeared.
How semantic Translation Memory works.
The fundamental difference between legacy TM and Flixu's approach is what a "match" means and what happens with it.
In a traditional CAT tool, TM match percentage measures character similarity — how many characters in the new string overlap with a previously translated string. A 70% match means 70% of the characters are the same. The tool substitutes the matched portion and leaves the rest for the translator. The AI (if any) fills in the gaps, but it doesn't know why that past translation was approved or what stylistic choices it reflected.
The Semantic Reranker identifies past translations that are conceptually and structurally similar to the current string — even when the exact words differ. "The vehicle came to a stop" and "The car halted" carry the same meaning; character-based matching would score them as low overlap. The Semantic Reranker recognizes them as equivalent for retrieval purposes.
Once a match is identified, it's used as a style reference — not a substitution. The language model receives the current string alongside the approved past translations as contextual examples. It generates a new translation that reflects the style and terminology of those references, rather than copying them directly. The output is organically consistent with your historical approvals, not mechanically copied from them.
The practical consequence: Even a small Translation Memory produces consistent output, because the model is learning your style from the references, not just matching strings. The Semantic Reranker works effectively even when a TM is small — the style context it provides improves output from the first approved translation.
How a single match improves the whole document.
In legacy TM workflows, each segment is translated in isolation. A match found in paragraph two helps paragraph two — it has no effect on paragraph twelve.
Flixu reads the entire document before translating any string. When the Semantic Reranker finds relevant past translations, those references are loaded into the translation context for the full document — not just the specific segments where they matched. A segment about "configuring the dashboard" in paragraph two informs the translation of "dashboard settings" in paragraph twelve, even if the strings don't match by character similarity.
This is why terminology consistency across a long document is a structural property of the approach, not a QA step that happens afterward.
Illustrative flow (conceptual)
Source document: 50-page technical manual (English)
↓
1. Semantic Reranker queries TM against full document
2. Finds conceptually relevant past translations
(engineering specs, UI terminology, formality register)
3. Past translations loaded as style references
4. Language model generates translation
with those references active throughout
5. Output: consistent terminology, matching register,
no drift between early and late sections
What the Translation Memory includes.
Semantic retrieval by meaning
Past translations are retrieved based on conceptual similarity, not character overlap. Strings that differ in exact wording but carry the same meaning surface as relevant matches. This expands the effective coverage of a TM — especially useful for early-stage workspaces where the TM is still building.
Reference injection, not substitution
Matched translations are provided to the language model as style references for the translation request — not substituted word-for-word. The model generates output that reflects the style and terminology of past approvals organically, without the awkward constructions that sometimes result from direct substitution in inflected languages.
Whole-document context
Relevant matches inform the translation of the entire document, not only the specific segments where they appeared. Terminology consistency across a long document is a structural outcome of the approach.
Post-Edit Learning Loop
When a team member confirms a corrected translation, it's added to the Translation Memory and available as a reference for subsequent requests. The TM improves with every approved correction — quick translations and full projects alike contribute to the shared pool.
TMX import and export
Existing Translation Memory from Trados, Phrase, MemoQ, or any other TMX-compatible platform imports directly. Your historical approved translations become available as style references from the first translation run in Flixu. Export is available in standard formats — no vendor lock-in.
When semantic TM changes the outcome.
SaaS teams translating product updates across sprints
A software product updated weekly generates new strings each sprint. Without Translation Memory, the same product term might be translated differently across sprints by different team members or MT runs. With semantic TM, each approved translation informs the next — "Dashboard" stays "Dashboard," formality stays consistent, and the UI reads as if it was localized by the same person across every release.
Teams using persistent TM across product update cycles typically find that terminology inconsistency — the same term appearing in multiple variants across the interface — drops from 15–25% of reviewed strings to under 2%.
→ SaaS localization workflows: Flixu for SaaS Teams
Agencies building a client-specific style corpus
An agency starting with a new client has no Translation Memory for them. Legacy TM would provide nothing useful until a significant volume of content is approved. The Semantic Reranker produces useful style references even from a small TM — ten approved segments carry enough stylistic signal to guide subsequent translations toward the client's approved voice.
As the agency works more for the client, the TM deepens. The longer the relationship, the more the output reflects the client's approved style without additional review overhead.
→ Agency workflows: Flixu for Agencies
Localization programs migrating from legacy TMS
Teams with years of approved translations in Trados, Phrase, or MemoQ can import their historical TM as a TMX file. The existing approved translations immediately populate the semantic retrieval layer — the first translation run benefits from the full accumulated approval history.
Frequently Asked Questions
How does Flixu handle fuzzy matches differently from traditional CAT tools?
+
Traditional CAT tools fill in the untranslated portion of a fuzzy match with machine translation, still processing the segment in isolation. Flixu uses the fuzzy match as a style reference for the full document context — the model generates output that reflects the tone and terminology of the match rather than substituting characters. A match in one paragraph improves consistency across the entire document.
What is the Semantic Reranker?
+
The Semantic Reranker is the retrieval component of Flixu's Translation Memory. It identifies past approved translations that are conceptually and structurally similar to the current string — understanding that "The vehicle stopped" and "The car halted" are equivalent for retrieval purposes, even though character overlap between them is low. This expands the effective coverage of a TM beyond what character-similarity matching would find.
Does Flixu support TMX file import from Trados, Phrase, or MemoQ?
+
Yes. Import a standard .tmx file from any TMX-compatible platform into any client profile. The imported translations are immediately available as semantic style references — your historical approvals begin informing new translations from the first run.
Does the Translation Memory update automatically after each translation?
+
It updates when a translation is confirmed by a team member (Translator, Project Manager, or Admin). If the output is accepted without modification, the session contributes to the TM. If the output is corrected before confirmation, the corrected version is saved — the Post-Edit Learning Loop captures the correction as the approved phrasing for future references.
Can I export my Translation Memory from Flixu?
+
Yes. Translation Memory exports in standard TMX format. There's no lock-in — your historical approved translations are yours to take to any TMX-compatible platform. Export is available from the workspace settings.
How does semantic similarity differ from fuzzy match percentage?
+
Traditional fuzzy match percentage uses edit distance (how many character insertions or deletions it takes to convert one string to another). A 70% fuzzy match means the strings are 70% similar by that measure. Flixu's semantic similarity measures conceptual closeness — two sentences that mean the same thing in different words can score as high similarity even when character overlap is low. This is what allows the Semantic Reranker to find useful references even when the exact wording has changed.
Does TM work across all 22+ supported languages?
+
Yes. The semantic retrieval and reference injection run across all supported language pairs. A Translation Memory built in English-to-German also informs English-to-French translations for the same client — the style references from one language pair contribute to the quality of others where the same stylistic patterns apply.
Connect your existing TM and see the difference.
Import your TMX file, run your first translation, and see how semantic retrieval compares to character-based matching on your own content.
Related Features
- LQA & Quality Assurance — Quality scoring that works alongside TM retrieval for auto-approval
- Team Collaboration — Shared TM across team members
- Brand Voice — Tone enforcement that complements semantic TM
- Client Management — Client-scoped TM isolation
- Translation Memory Guide — Educational deep-dive on how TM works