SuperMemory is the AI-maintained translation knowledge base built into Supervertaler for Trados. Inspired by Andrej Karpathy's "LLM Knowledge Base" pattern (April 2026), it complements translation memories and termbases — which store wordings and term pairs — with a structured Markdown wiki holding the reasoning behind translation decisions, which the AI processes, maintains and consults.
Knowledge lives in a memory bank: an Obsidian-compatible folder of interlinked Markdown articles.
This repository is the format definition, NOT a memory bank. It holds SPEC.md and a starting skeleton/. Real banks live under the user's data folder (<user data>\memory-banks\<name>) and must never be kept inside a clone of this repo — a live bank sat here behind a symlink until 2026-07-28, which left real client data one git add -A away from a public push. The skeleton is nested under skeleton/ specifically so that cannot recur.
00_INBOX/ — Raw material drop zone (articles, glossaries, feedback, style guides)
01_CLIENTS/ — Client profiles (preferences, style rules, terminology decisions)
02_TERMINOLOGY/ — Term articles with approved translations, rejected alternatives, and reasoning
03_DOMAINS/ — Domain knowledge (legal, medical, tech, marketing — conventions and pitfalls)
04_STYLE/ — Style guides and formatting conventions
05_INDICES/ — Auto-generated indexes and maps of content
06_TEMPLATES/ — AI agent prompt templates (Process Inbox, Health Check, query, translate)
- Ingest: User drops raw material into
00_INBOX/ - Process Inbox: AI reads raw material, produces structured articles in 01–04, creates
[[backlinks]] - Health Check: AI periodically scans the vault for inconsistencies, broken links, stale content, and missing cross-references
compile.md— Process Inbox: raw material → structured articleslint.md— Health Check: scans KB for problems, auto-fixes what it cantranslate_with_kb.md— Translation agent: consults KB before translatingquery.md— Query agent: answers questions from KB content
The following terms are used consistently across the plugin UI, help docs, and vault templates:
| User-facing term | Internal / file name | Description |
|---|---|---|
| Process Inbox | compile.md |
Reads raw inbox files, writes structured articles |
| Health Check | lint.md |
Scans vault for problems, fixes and reports |
| Quick Add (Ctrl+Alt+M) | SuperMemoryQuickAddAction |
Captures terms from Trados editor |
| Active prompt | ActivePromptPath |
Per-project prompt for Quick Add and Batch Translate |
Avoid using "compile", "compilation agent", "lint", or "linting agent" in user-facing text. These are internal terms only.
Primary integration: Supervertaler for Trados (C#/.NET plugin, C:\Dev\Sv\Supervertaler-for-Trados\)
Future: Supervertaler Workbench (Python), Supervertaler Workbench v2 (Tauri/Rust)
- GitHub repo: Ships the skeleton (folders,
_EXAMPLE_*files, templates) - User data folder:
C:\Users\{user}\Supervertaler\supermemory\ - Developer data: Real translation data lives in the deployed vault, excluded from Git by
.gitignore - Pattern matches existing deployment: prompt library at
C:\Users\{user}\Supervertaler\prompt_library\
_EXAMPLE_*files are tracked and shipped — they show users the format06_TEMPLATES/is fully tracked — agent prompts are shipped- All other
.mdfiles in 00–05 folders are gitignored (user data) - The
.obsidian/folder is partially tracked (core config shipped, workspace/cache ignored)
- No vector DB, no RAG. Structured Markdown + LLM reasoning is sufficient at translation project scale.
[[Backlinks]]are mandatory. They make the vault navigable in Obsidian's graph view and allow the LLM to follow connections.- Term articles record WHY. Rejected alternatives and reasoning prevent re-litigation of terminology decisions.
- Client overrides trump general rules. The terminology hierarchy: client profile > term article > domain conventions > general style guide.