SuperMemory — the Supervertaler memory-bank format

SuperMemory is the knowledge-base system built into Supervertaler for Trados. Where a translation memory stores previous wordings and a termbase stores approved term pairs, SuperMemory stores the reasoning: why a term was chosen, what a client insists on, what was rejected last time, and which domain conventions apply.

That knowledge lives in a memory bank — a plain folder of Obsidian-compatible Markdown articles with YAML frontmatter and [[backlinks]]. No database, no embeddings, no vendor lock-in. If every tool in this ecosystem disappeared tomorrow, your knowledge would still open in any text editor.

This repository is the format definition: the specification, and a starting skeleton.

What's here

Path	What it is
SPEC.md	The format specification (v1.1) — folder layout, frontmatter schema, ### FILE: output markers, scoring rules, code-fence tolerance
skeleton/	A ready-to-copy empty bank: the seven folders, .obsidian/ defaults, one _EXAMPLE_ article per folder, and the agent prompt templates

This repository is not itself a memory bank. The skeleton is deliberately nested under skeleton/ so that no one is tempted to point a live bank at a clone of this repo. Keep your real banks well away from version control you might publish — or if you do want history, use a repository with no remote.

Where banks actually live

Supervertaler for Trados keeps banks under your user-data folder:

<user data>\memory-banks\
├── default\
├── acme-legal\
└── pharma\

Each is a self-contained bank with the seven-folder skeleton. You switch between them from the Memory Bank dropdown in the Supervertaler Assistant panel. The plugin ships its own copies of the templates and writes the skeleton for you whenever you create a bank, so you never need to clone this repo to get started — it is here as the reference, not as a dependency.

Reading a bank from outside Trados

Since v18.20.146 the plugin exposes memory banks over the Supervertaler MCP server, so any MCP client (Claude Desktop, Claude Code, and others) can consult your bank while you work — whatever CAT tool you happen to have open:

• get_supermemory_context — the bank for the current project, with the article paths it drew from
• search_supermemory — free-text search across the active bank
• list_supermemory_banks — which banks exist and which is active

Because a bank is just Markdown on disk, you can also open it in Obsidian, search it with ordinary tools, or point a filesystem MCP server at the folder.

The seven folders

Folder	Contents
00_INBOX	Raw material awaiting processing — briefs, feedback notes, reference articles
01_CLIENTS	Client profiles: preferences, style rules, terminology decisions
02_TERMINOLOGY	Term articles with approved translations, rejected alternatives, and the reasoning
03_DOMAINS	Domain conventions and common pitfalls
04_STYLE	Style guides, register notes, formatting rules
05_INDICES	Auto-generated indexes
06_TEMPLATES	The agent prompts that drive Process Inbox, Health Check and Distill

Full details, including the frontmatter schema and the rules a conformant host must follow, are in SPEC.md.

Design notes

The approach is inspired by Andrej Karpathy's LLM knowledge base pattern: structured Markdown that a model reasons over directly, rather than embeddings and retrieval. At translation-project scale — hundreds of articles, not millions — that stays auditable in a way a vector store does not. Every answer traces to a file you can open, read and correct by hand.

History

An earlier plan for a standalone cross-platform "Supervertaler Assistant" app that would read these banks outside Trados was not pursued; the MCP server covers that ground without a second application to maintain. Its code remains in supervertaler-assistant for reference.

Licence

MIT — see LICENSE.