↳ no match yet: translate once, then it’s remembered
✓Cancelin memory
↳ from memory: «Annuler»
✓Settingsin memory
↳ from memory: «Paramètres»
+Export as PDFnew
↳ no match yet: translate once, then it’s remembered
✓Sign outin memory
↳ from memory: «Se déconnecter»
+Delete accountnew
↳ no match yet: translate once, then it’s remembered
5 of 8 lines were already in memory. You only translate the 3 that are new.
Most of what you translate, you’ve translated before, so those lines come back with what the translation memory already holds. You only pay for the new ones.
传统的翻译工具曾利用它来节省时间。而在 AI 时代,它的作用愈发关键:它能向模型展示您的翻译标准,明确预期的翻译决策。其成效极其显著:
代理与人类之间的讨论背景:包括聊天记录、回复串、搜索历史以及 Slack 或 Teams 中的上下文信息。
Figure 2 · Anatomy of a TM record
What a translation memory stores now
just the pair
SourcePure steel rejects lacquers and paints
Target«Чиста сталь відкидає лаки та фарби»
Metadataapproved by · date · file & context
…and what the agents also kept
Thinking trace
weighing "rejects lacquers and paints"
├ literal «відкидає лаки та фарби» keeps the metaphor ✓
├ smoother «не приймає покриття» clearer, but flattens it
└ decision go literal, the bluntness is the point
Glossary
steel → сталь🔒paints → фарби
Handoff note
translator → editor
“Kept the steel metaphor literal. Check it reads naturally in UK.”
Discussion
Mariais «відкидає» too harsh for a product line?
LiteressIt mirrors the source’s bluntness. I’d keep it.
Switch to AI-era to see what an agentic pipeline adds →
Classic CAT memory kept three things: the source, the target, and a little metadata. Toggle to AI-era to see what an agentic pipeline also keeps: the glossary it locked, the note it handed off, the reasoning it went through, the conversation behind the call.
这些信息对人类依然有用,但对 AI 来说却至关重要。它让模型能够重构出人类译者在文档、大脑和对话中承载的上下文。讽刺的是,大语言模型驱动的翻译比人类更需要 TM 和 CAT 工具——毕竟,人类只需纸、笔和字典就能顺利完成翻译。
Exact hit. The TM has translated this verbatim before, so the system can fill “Enregistrer les modifications” automatically.
Pick a phrase to translate. Exact matches auto-fill, fuzzy matches flag for review, and semantic matches surface meaning even when no words overlap. Real systems run all three at once.
In general-domain space, “ileum” and “ilium” collapse together
colorgutbone
ileum ↔ ilium
8units · collapsed
crimson ↔ scarlet
10units · true synonyms
For a general-domain embedder, the gap between ileum (intestine) and ilium (hip bone) is as small as between crimson and scarlet: a one-letter typo the model reads as a synonym. Toggle to a domain-aware setup and the medical guardrail pulls them apart. This is why 2026 TM needs reranking, not raw vectors.
因此,护栏机制必不可少。到 2026 年,翻译记忆库的默认配置将包括:
混合搜索:整合精确匹配、模糊匹配和语义匹配,从过往翻译历史中筛选出最相关的结果。
重排序:根据实际上下文对候选译文的相关性进行重新评分。
树遍历:人工或智能体可以从某个匹配项跳转到其最近邻,从而对数据库进行深入探索。
Figure 5 · The 2026 default stack
Retrieve wide, then rerank narrow
Query
A new source segment arrives. No assumptions yet about which kind of memory will help.
Exactidle
Fuzzyidle
Semanticidle
Candidates
Waiting for retrieval…
Step 1 / 5
Modern TM is not one search; it’s parallel exact + fuzzy + semantic recall, deduped, then reranked against the real context, with tree traversal to explore. Raw vector recall alone would have kept “Reset device”; reranking drops it.
从这个角度看,市场可以划分为三个层级:记忆即复用、记忆即治理、以及记忆即 AI 燃料。Transept 则押注于第四个层级:记忆即决策上下文。
Figure 6 · The market, four levels deep
What each tool actually remembers
Tool
What it remembers
AI → memory
Tool
What it remembers
AI → memory
Level 1 · Memory as reuse“Have we translated this before?”
Trados / RWS
Have we translated this before?
Your past segments, with exact, fuzzy, and in-context matches.
AI is bolted on around it; at heart it is still segment reuse.
memoQ
Is this the same segment, in the same spot?
Segments plus where they sit (its 101% / 102% matches check the surroundings).
Clever retrieval, but the thing remembered is still the segment.
Wordfast
Have we translated this before?
Shared TMs, glossaries, and QA you can reach from a browser.
Built for access and reuse, not for remembering decisions.
OmegaT / CafeTran
Have we translated this before?
Open-source fuzzy matching across several TMs and glossaries.
Proves good reuse is not enterprise-only; still no decision trail.
Level 2 · Memory as governance“Which memory do we trust?”
Phrase TMS
Which memory do we trust here?
TM, term bases, MT profiles, and QA under one roof.
The TM still just holds reusable segments for pre-translation.
Crowdin
Which memory do we trust here?
Per-project TMs; it can keep only the approved translations.
Remembers the approved text, not the reasoning behind it.
Smartcat
Whose memory is this?
Memory sorted by client, team, and workspace.
Auto-attaches the right TMs; one writable, the rest read-only.
XTM Cloud
Does this translation deserve to be memory?
Approved vs unapproved entries; raw MT is not saved by default.
Trust control: a setting decides if unapproved memory can be suggested.
Wordbee
Live work, or durable memory?
A master TM plus temporary per-project memories.
The good bits get promoted into the master TM, by hand or automatically.
Bureau Works
Who is allowed to write to memory?
Memories tied to departments, with read / write permissions.
Blends TM, LLM, MT, and glossary into one stream of suggestions.
MateCat
Can corrections feed the engine?
Public / private MyMemory; live edits improve the MT as you go.
Feeds adaptive MT, but memory still means segments and fixes.
Level 3 · Memory as AI fuel“Can AI output become memory?”
Lilt
Can memory train the engine?
Confirmed pairs and terms that sharpen its predictions over time.
Tunes the model itself; keeps no record of the rejected options.
Smartling
Is this memory human or machine?
AI output in its own store; human work stays in the regular TM.
Clear provenance: machine memory never poses as human-approved.
Lokalise
Did a human bless this AI output?
AI translations enter the TM once a reviewer accepts them.
Human-gated; safe, but slow for small teams in the trenches.
Transifex
Is the AI output good enough to keep?
A quality score decides if AI output auto-enters the TM.
Hands-off, as long as you trust the proprietary score.
Phrase Language AI
Which engine and memory for this job?
Routing across engines, with quality estimates and glossaries.
The TM is just one input beside MT, glossary, and QA.
Level 4 · Memory as decision context“Why did this translation win?”
TranseptOURS
Why did this translation win?
The segment plus the rejected drafts, the discussion, the reasoning, QA, and version history.
Memory is decision context, shared by people and agents and fed back to the model.
Level 1 · Memory as reuse“Have we translated this before?”
Trados / RWS
Have we translated this before?
Remembers: Your past segments, with exact, fuzzy, and in-context matches.
AI → memory: AI is bolted on around it; at heart it is still segment reuse.
memoQ
Is this the same segment, in the same spot?
Remembers: Segments plus where they sit (its 101% / 102% matches check the surroundings).
AI → memory: Clever retrieval, but the thing remembered is still the segment.
Wordfast
Have we translated this before?
Remembers: Shared TMs, glossaries, and QA you can reach from a browser.
AI → memory: Built for access and reuse, not for remembering decisions.
OmegaT / CafeTran
Have we translated this before?
Remembers: Open-source fuzzy matching across several TMs and glossaries.
AI → memory: Proves good reuse is not enterprise-only; still no decision trail.
Level 2 · Memory as governance“Which memory do we trust?”
Phrase TMS
Which memory do we trust here?
Remembers: TM, term bases, MT profiles, and QA under one roof.
AI → memory: The TM still just holds reusable segments for pre-translation.
Crowdin
Which memory do we trust here?
Remembers: Per-project TMs; it can keep only the approved translations.
AI → memory: Remembers the approved text, not the reasoning behind it.
Smartcat
Whose memory is this?
Remembers: Memory sorted by client, team, and workspace.
AI → memory: Auto-attaches the right TMs; one writable, the rest read-only.
XTM Cloud
Does this translation deserve to be memory?
Remembers: Approved vs unapproved entries; raw MT is not saved by default.
AI → memory: Trust control: a setting decides if unapproved memory can be suggested.
Wordbee
Live work, or durable memory?
Remembers: A master TM plus temporary per-project memories.
AI → memory: The good bits get promoted into the master TM, by hand or automatically.
Bureau Works
Who is allowed to write to memory?
Remembers: Memories tied to departments, with read / write permissions.
AI → memory: Blends TM, LLM, MT, and glossary into one stream of suggestions.
MateCat
Can corrections feed the engine?
Remembers: Public / private MyMemory; live edits improve the MT as you go.
AI → memory: Feeds adaptive MT, but memory still means segments and fixes.
Level 3 · Memory as AI fuel“Can AI output become memory?”
Lilt
Can memory train the engine?
Remembers: Confirmed pairs and terms that sharpen its predictions over time.
AI → memory: Tunes the model itself; keeps no record of the rejected options.
Smartling
Is this memory human or machine?
Remembers: AI output in its own store; human work stays in the regular TM.
AI → memory: Clear provenance: machine memory never poses as human-approved.
Lokalise
Did a human bless this AI output?
Remembers: AI translations enter the TM once a reviewer accepts them.
AI → memory: Human-gated; safe, but slow for small teams in the trenches.
Transifex
Is the AI output good enough to keep?
Remembers: A quality score decides if AI output auto-enters the TM.
AI → memory: Hands-off, as long as you trust the proprietary score.
Phrase Language AI
Which engine and memory for this job?
Remembers: Routing across engines, with quality estimates and glossaries.
AI → memory: The TM is just one input beside MT, glossary, and QA.
Level 4 · Memory as decision context“Why did this translation win?”
TranseptOURS
Why did this translation win?
Remembers: The segment plus the rejected drafts, the discussion, the reasoning, QA, and version history.
AI → memory: Memory is decision context, shared by people and agents and fed back to the model.
Everyone can store the final translation. The real differences are what the system remembers, when it trusts that memory, and where it uses it next. Filter by level; Transept’s bet is level 4: memory as decision context.
这正是我们在 Transept 押下的赌注。无论句子本身多么重要,翻译记忆库都不应只记住句子。它应该记住那些让句子变得可靠的工作。TM 必须成为人类和 AI 共同享有的决策上下文。
Figure 7 · Memory as decision context
Transept remembers the work behind the sentence
The approved segment · EN → FR · idiom
It costs an arm and a leg
«Ça coûte les yeux de la tête»
Everyone stores this
Context fed to the next translationkept 1 / 7
✓«Ça coûte les yeux de la tête»· the French idiom, same idea and register
✗«Ça coûte un bras et une jambe»· a literal calque, reads as a translation
✗«C’est très cher»· accurate, but flattens the colour
"costs an arm and a leg": an idiom, translate the meaning
· a literal calque would read as a translation
· plain "très cher" is accurate but loses the colour
→ use the French idiom for the same idea
✓idiom mapped, not calqued
✓register matches source (casual)
⚠length checked, fits the button
v1machine«un bras et une jambe»
v2Literess«les yeux de la tête»
v3humanapproved · current ✓
MariaKeep the body-part image, or go fully idiomatic?
LiteressFrench has its own: «les yeux de la tête». Same register, lands natively.
🔍web: “arm and a leg french equivalent”📖dictionary: coûter les yeux de la tête💬comment: matches our playful brand voice
Right now the next translation sees only the final sentence, the same memory the market keeps. It can copy the style; it can’t replay the logic.
Translation generates engineering gold: the rejected drafts, the reasoning, the review trail. Almost everyone throws it away. Switch on the layers we keep behind one idiom, and watch how much more the next translation gets to see.
TM sources · what feeds each agent✓ kept · ✗ dropped, with the reason
This documentthis run · live
●Ajoutez un composant.live draft, used for consistency
Team / project docsshared, approved
✓dashboard → tableau de bord
✗«maquette en cours»dropped: WIP, no editorial sign-off
Org librarywhole org, broadest
✓Sign in → Se connecter
✗«se loguer»dropped: rejected alternative
Glossary & styleguidelocked rules
🔒widget → composant
🔒formal «vous»
↓
every agent grounds on the kept sources, plus its own live drafts
TranslatingAgent A¶ 1–14k
Add a widget.→ Ajoutez un composant.🔒 locks widget → composant, syncs →
TranslatingAgent B¶ 14–27k
Remove the widget.→ Supprimez le composant.✓ uses synced composant
TranslatingAgent C¶ 27–40k
Configure the widget.→ Configurez le composant.✓ uses synced composant
Phase 1. Each agent reads the kept sources and its own live drafts. The instant Agent A locks widget → composant it syncs to the glossary, so all three land on composant without waiting on each other.
8h→50minfor a 40,000-word document, consistent end to end.
A big document is translated by many agents in parallel, then proofread in parallel: the same job at the same time, never mixed. Each agent grounds on a graded stack of sources, and not everything makes the cut.
扩展语境与工作流控制
在掌握了捕捉优秀决策与人类才智的方法后,我们进一步深挖,致力于提供深层语境支持与便捷的工作流:
全渠道语境:通过聊天、评论和文档搜索,人类译员和 AI 都能找寻翻译背后的深层背景——包括网页搜索、词典查询、编辑器评论,以及与队友和 Literess 的讨论。