Voice notes. Bookmarks. Tweets. YouTube transcripts. All captured automatically, filed as plain markdown, queryable from anywhere.
This is how the pieces fit together.
The architecture
┌── Telegram bots ──────────────┐
│ URL bot ← bookmarks, │
│ articles, │
│ YouTube │
│ Voice bot ← voice memos │
│ Agent bot ← queries │
└───────────────┬───────────────┘
↓
┌──────── n8n ────────┐
│ Bookmark Summary │
│ Voice Notes → Vault│ ← nervous system
│ Vault Writer │
│ Weekly Promote │
└──────────┬──────────┘
↓
┌─── Vault on VPS ────┐
│ wiki/sources/ │ ← source of truth
│ wiki/entities/ │ plain markdown
│ wiki/concepts/ │
│ wiki/learnings/ │
└──────────┬──────────┘
↕ Syncthing (~10s)
┌──────────┴──────────┐
│ Mac vault │ ← Obsidian opens here
└──────────┬──────────┘
↑
┌──────────┴──────────┐
│ Agent (read-only) │ ← answers
│ vault-aware │ writer│ via Telegram
└─────────────────────┘What it does
- Capture is automated. Three Telegram bots cover all input. Share-sheet a URL, hold-to-record a voice memo, ask the agent a question — the right workflow handles it.
- The vault lives on a VPS. Plain markdown on disk. Source of truth is the server, so write workflows don’t depend on the laptop being open.
- Sync is real-time. Syncthing mirrors VPS ↔ Mac in about ten seconds. Open Obsidian on the Mac, see what got captured five minutes ago.
- The agent reads the whole thing. From the same phone that captures notes, you can query them. The agent is read-only by default; drafts go to a separate inbox folder for manual review.
- The wiki grows itself. A Sunday cron job reads the week’s captures, extracts entities and concepts and learnings, and writes new wiki pages with
[[wikilinks]]back to sources.
The vault schema
vault/
├── wiki/
│ ├── index.md ← topic-led entry point
│ ├── log.md ← every wiki op
│ ├── decisions.md ← dated decision log
│ ├── roadmap.md ← open work
│ ├── sources/
│ │ ├── bookmarks/ ← URL bot writes here
│ │ └── voice-notes/ ← voice bot writes here
│ ├── entities/ ← people, companies, products
│ ├── concepts/ ← ideas, frameworks
│ ├── learnings/ ← takeaways
│ └── lists/ ← /todos /projects /other
└── raw/ ← immutable source dumpsOne home per fact. No duplicates.
The components
| Layer | Tool | What it does |
|---|---|---|
| Capture (URLs) | Telegram bot | tweet · article · YouTube → markdown |
| Capture (voice) | Telegram bot | voice → Whisper → categorized markdown |
| Routing | n8n | webhooks, schedules, transforms |
| Storage | VPS | plain markdown, schema-enforced |
| Sync | Syncthing | VPS ↔ Mac, ~10s |
| Search | QMD | semantic search inside agent container |
| Agent | NousResearch Hermes | read-only vault, drafts via inbox |
| Promotion | weekly cron | bookmarks → entities/concepts |
Every piece is open source.
Three patterns
One home per fact.
WHY (decisions, roadmap, learnings) lives in the vault. HOW (operational state) lives in per-project CLAUDE.md. About YOU (preferences, env quirks) lives in agent memory. WHAT (code) lives in git. If a fact appears in two files, one of them links to the other. Never copy.
Bots write, humans browse.
Capture is automated. Reading is human. Telegram is the lowest-friction surface available — share-sheet on phone, hold-to-record for voice. The bot replies with a confirmation and a delete button so a wrong capture is one tap to remove. Reading still happens in Obsidian, where it should.
Promote weekly.
Every Sunday, n8n runs a “promote” job: it reads new bookmarks, asks the model to extract entities and concepts and learnings, and writes them as new wiki pages with [[wikilinks]] back to sources. The wiki grows itself without anyone opening it.
The agent loop
The on-VPS agent has two personalities:
- vault-aware — read-only. Queries via grep and semantic search. “What did I save about Karpathy’s LLM wiki idea?”
- vault-writer — drafts new pages to a
vault-inbox/folder. The vault itself stays read-only. Drafts get promoted manually after review.
The default model is a fast cheap one (Grok 4.1 Fast at the moment, ~15× cheaper than Sonnet for this workload). There’s an escape hatch to Sonnet for harder queries.
What this isn’t
- Not a content production system. This is for thinking.
- Not turnkey. Six pieces have to talk to each other; getting them aligned takes a weekend.
- Not novel pieces. Every component is open source. The recipe is what’s rare.
Take it
There’s a template repo with the architecture, schema, and skill examples. Star it if you’re building something similar — that’s the signal to ship more (workflow JSONs, bot code, deploy templates).
Follow @arthurpfz for build logs.