SuperMemory is a new feature in Supervertaler that replaces the traditional “glossary + translation memory” approach with something more powerful: a living, AI-maintained knowledge base that understands why you translate things the way you do.
The problem with traditional TM/TB
We’ve all been there. You open a project for a client you haven’t worked for in six months. You know you made specific terminology decisions back then — but why did you choose “naleving” over “conformiteit”? What was the client’s preference on register? Was there a style guide somewhere?
Traditional termbases give you flat term pairs. Translation memories give you previous translations without context. Neither captures the reasoning, the client preferences, or the domain conventions that informed your decisions.
What SuperMemory does differently
SuperMemory stores your translation knowledge as interlinked Markdown articles in an Obsidian vault — human-readable, portable, and organised into five categories:
- Client profiles — language preferences, terminology decisions, style rules, project history
- Terminology articles — approved translations with rejected alternatives and the reasoning behind each choice
- Domain knowledge — conventions, common pitfalls, and reference material for specific fields
- Style guides — formatting rules, register, localisation conventions
- Auto-generated indexes — the AI maintains these to keep everything connected
The articles are interlinked with backlinks, so you can navigate from a client to their preferred terms to the domain those terms belong to. Open the vault in Obsidian and you get a visual knowledge graph showing all the connections.
The workflow
1. Ingest — Drop raw material into the inbox: client briefs, style guides, glossaries, feedback notes, reference articles. You can also use the Obsidian Web Clipper to clip web pages straight into your inbox from any browser.
2. Process Inbox — Click the button in the Supervertaler Assistant toolbar. The AI reads your raw material and writes structured knowledge base articles, filing them into the right folders with proper backlinks and metadata.
3. Health Check — Click the second button. The AI scans your entire vault for problems: conflicting terminology, broken links, orphaned articles, stale content, missing cross-references. It fixes what it can and flags the rest for your review.
4. Translate — This is where it gets interesting. When you translate (batch or single segment), Supervertaler automatically loads the relevant KB articles into the AI prompt. It detects the client from your project name, identifies the domain from the document content, and pulls in the matching client profile, domain knowledge, style guide, and terminology articles — complete with the reasoning behind each decision.
The AI doesn’t just know what to translate “compliance” as — it knows why you chose “naleving” for this specific client, that the client rejected “conformiteit” last year, and that in this legal domain you should preserve the passive voice.
How it works with existing tools
SuperMemory doesn’t replace your termbases and TMs — it adds an intelligence layer on top. Your MultiTerm glossaries still provide the flat term pairs. Your TMs still provide style anchoring from previous translations. SuperMemory adds the why: the reasoning, the client context, the domain conventions, and the style rules that make the difference between a correct translation and a great one.
Technical details
- Built on plain Markdown files — no database, no vendor lock-in, works with any text editor
- Inspired by Andrej Karpathy’s LLM Knowledge Base architecture
- No RAG, no vector database, no embeddings — at translation project scale, structured Markdown with an LLM acting as your librarian outperforms retrieval-based approaches
- Open source: github.com/Supervertaler/Supervertaler-SuperMemory
Try it
SuperMemory is available now in Supervertaler for Trados. The vault ships as a skeleton with example files — drop your first client brief into the inbox and click Process Inbox to see it in action.
Documentation: supervertaler.gitbook.io/trados/features/supermemory
I’d love to hear your thoughts — especially if you have ideas for what kind of knowledge you’d want to capture beyond terminology and client profiles.