A July 2026 post from chewa. claimed "ANTHROPIC'S LEAD ENGINEER ACCIDENTALLY LEAKED HIS PERSONAL OBSIDIAN VAULT" — 8,893 nodes, 4,729 links, overlay captions comparing the graph to 21 inputs, hidden layers, ReLU activation, and a $10-a-month app running decisions inside the company. The clip crossed 1.1M views on X.
X Community Notes rated the central claim misleading: the quoted article describes a 33-year-old science journal editor in Porto — not an Anthropic engineer. Replies called it engagement farming and noted a knowledge graph does not run or decide anything — it is a static visualization of links between notes.
This guide does two jobs. First, it fact-checks the viral frame so you do not rebuild the wrong story. Second, it explains what an Obsidian vault actually is and how developers build self-writing vaults that agents maintain — the real pattern behind the hype, without the neural-network cosplay.
TL;DR — Questions the viral post raised
| Question | Honest answer |
|---|---|
| Was this an Anthropic leak? | No — Community Notes tie the source to a Porto journal editor's personal vault, not Anthropic |
| Is the graph a neural net? | No — it is Obsidian Graph View over [[wikilinks]]; static, not executing |
| Do 8,893 nodes "fire in real time"? | No — nodes are markdown files; the animation is UI, not inference |
| What is an Obsidian vault? | A folder of .md files + the Obsidian app to browse, link, and search them |
| What makes vaults "self-writing"? | Agents + conventions — CLAUDE.md, folder rules, cron or /loop, ingest/query/lint |
| Cost? | Free locally; ~$10/mo if you want Obsidian Sync |
| Same app Anthropic uses internally? | Unverified — Anthropic may use markdown knowledge bases; this viral post is not proof |
The viral clip (what you are actually seeing)
Screen recording from the chewa. thread — Obsidian's graph camera panning across a dense link structure:
Demo clip (WebM): the viral video shows Obsidian Graph View — each dot is a note, each line is a [[wikilink]]. Impressive PKM hygiene; not a forward pass through hidden layers.
What the overlay text implies vs what is on screen:
| Viral claim | What is actually happening |
|---|---|
| "21 inputs, 10+ hidden layers, ReLU" | Graph layout algorithm positioning nodes — no activation function |
| "Connections firing in real time" | CSS/WebGL animation as you pan and zoom |
| "Living brain that runs decisions" | You (or an agent you configure) read notes and decide; the graph does not |
| "$2M/year for markdown in folders" | Discipline + retrieval quality can be valuable; the salary figure is unverified hype |
| "Anthropic's internal knowledge system" | Debunked by Community Notes on the source article |
What is an Obsidian vault?
An Obsidian vault is simply a directory on your filesystem that Obsidian treats as one notebook.
Core anatomy
my-vault/
├── .obsidian/ # App settings, plugins (optional in git)
├── Daily/ # Your folder convention — you choose
├── Projects/
├── People/
├── inbox.md
└── index.md # Optional map-of-content (MOC)
Every note is a .md file. Links use [[wikilink]] syntax or standard markdown [label](path.md). Attachments sit beside notes or in an attachments/ folder. Because files are plain text:
- Git tracks history
- Claude Code reads and edits directly
- Any editor works if Obsidian disappears tomorrow
Obsidian adds search, graph view, backlinks pane, plugins (Templater, Dataview, Tasks), and optional Sync / Publish. The vault is yours; the app is the lens.
Vault vs Notion vs Files.md
| Obsidian vault | Notion | Files.md | |
|---|---|---|---|
| Storage | Local .md files | Cloud database | Local .md in browser |
| Links | [[wikilinks]] + graph | Relational DB links | Minimal structure |
| Agent-friendly | Excellent — paths + git | API required | Excellent — by design |
| Philosophy | Build your system | Team wiki / DB | Radical simplicity |
See Files.md: local-first note-taking for the anti-complexity pole; Obsidian sits in the middle — powerful, but you own the scaffolding cost.
Graph view is a map — not a mind
Obsidian Graph View renders notes as nodes and [[wikilinks]] as edges. Filters can hide orphans, tag subsets, or color by folder.
What it helps with:
- Spot orphan notes with no links
- See clusters (projects, people, concepts)
- Navigate serendipitous connections while writing
What it does not do:
- Run embeddings search (unless you add plugins)
- Execute agent tools
- Decide anything — no forward pass, no weights, no training
A vault with 8,893 notes and 4,729 links is a large, well-linked PKM corpus — impressive maintenance — not a replacement for Claude or a company decision engine. Critics on the viral thread (David Dimlich and others) made exactly this point: "A knowledge graph doesn't run or decide anything. It's a static visualization."
The real pattern: self-writing vaults + agents
The useful idea buried in the viral post is older than the neural-network overlay: point an agent at a markdown vault and let it maintain structure while you capture raw thought.
Karpathy LLM Wiki (the intellectual source)
Andrej Karpathy's LLM Wiki gist defines three layers:
- Raw sources — immutable drops (articles, transcripts, clips)
- Wiki layer — agent-maintained summaries, entity pages, synthesis
- Schema —
CLAUDE.md/AGENTS.mddefining ingest, query, lint
Operations: ingest (new source → update 10–15 linked pages), query (read wiki, cite), lint (contradictions, stale claims, orphans). explainx.ai's Karpathy LLM Wiki guide walks the full pattern.
Google's OKF — enterprise formalization
Open Knowledge Format (OKF) (June 2026) standardizes the same shape: markdown + YAML frontmatter, index.md, log.md, typed concept pages. Karpathy's gist is the idea; OKF is the interchange spec.
Chewa's "Self-Writing Vault" frame (Jun 29, 2026)
The article chewa. quoted — "The Self-Writing Vault: 8 Rules for Pointing Claude at Obsidian and Letting It Run Without You" — describes:
- One evening of setup
- A vault that files itself at 7 a.m.
- A model that opens every session already knowing your work
That maps to mechanics builders already ship:
| Rule (conceptual) | Implementation |
|---|---|
| Folder semantics | 00-inbox/, 10-raw/, 20-wiki/, 30-pdca/ (pdca-llm-wiki) |
| Agent brief | Root CLAUDE.md loaded every Claude Code session |
| Scheduled maintenance | /loop 24h or cron calling ingest + lint |
| Machine-readable notes | Frontmatter, [[wikilinks]], "for future Claude" preambles |
| Separation raw vs wiki | Never mutate sources; synthesize in 20-wiki/ |
| Verification | Lint pass for broken links, stale dates, contradictions |
| Session continuity | Hooks on SessionStart / SessionEnd (Claude's Journal pattern) |
| Start tonight | Clone a template vault; add 10 notes; run first ingest |
None of this requires Obsidian-specific infrastructure — the vault is a folder. Obsidian is optional UI on top.
Minimal starter: your first agent-ready vault tonight
Step 1 — Create the folder
mkdir -p ~/vault/{00-inbox,10-raw,20-wiki,90-system}
cd ~/vault && git init
Open ~/vault in Obsidian (File → Open folder as vault). Free tier is enough.
Step 2 — Add CLAUDE.md
# Vault agent brief
## Folders
- `00-inbox/` — quick captures, unprocessed
- `10-raw/` — immutable sources (PDF excerpts, transcripts)
- `20-wiki/` — synthesized entity and concept pages
- `90-system/` — rules and index
## On ingest
1. Read new file in `00-inbox/` or `10-raw/`
2. Update `20-wiki/` entity pages and `index.md`
3. Append one line to `90-system/log.md`
## On query
Search `index.md` first, then drill into linked wiki pages. Cite paths.
## On lint
Report orphan wiki pages, broken [[links]], claims older than 90 days without refresh.
Step 3 — Wire Claude Code
cd ~/vault
claude
Claude Code reads CLAUDE.md automatically when cwd is the vault. For scheduled filing:
/loop 24h run vault lint, process 00-inbox, update index.md
See Claude Code loops official guide for /loop vs /goal vs turn-based patterns.
Step 4 — Optional Obsidian plugins
| Plugin | Why |
|---|---|
| Templater | Daily note templates with frontmatter |
| Periodic Notes | Journal cadence |
| Local REST API | HTTP read/write for external agents |
| Tasks | Checkbox queries across vault |
Step 5 — Connect to agent memory elsewhere
Vault notes complement — not replace — repo-level MEMORY.md and Karpathy-style wikis. Code projects keep CLAUDE.md; life and research keep the Obsidian folder.
Open-source vault templates worth cloning
| Project | What you get |
|---|---|
| gyeorye/pdca-llm-wiki | Drop-in vault + PDCA retrospectives + rule files |
| obsidian-second-brain | Self-rewriting vault skill, AI-first note rules |
| jgcosme/claude-obsidian-memory | Claude Code plugin — hooks + vault-search skill |
| dys5315/domvault | Self-synthesis loop + optional Constellation publish |
| cast-claudes_journal | Per-date journal entries as graph nodes |
These repos prove the pattern is markdown discipline + agent contracts — not a proprietary neural appliance.
What people got wrong (and right) in the replies
Wrong:
- Treating graph animation as ML inference
- Assuming Anthropic internal ops from an unverified screenshot
- Expecting a $10/mo app to replace agent harness engineering
Right:
- Large linked vaults are genuinely hard to maintain — automation helps
- Markdown + agents is how many builders store memory in 2026
- Discipline over infrastructure — chewa.'s quoted piece emphasizes three years of filing, not a custom server
@ProjectorBach — "Every 'leak' is a marketing maneuver" — is the right prior when a post pairs a sensational employer with a product screenshot and no primary source.
Obsidian vault vs RAG vs MEMORY.md
| Approach | Best when |
|---|---|
| Obsidian vault (wiki pattern) | Personal knowledge, cross-linked concepts, human browse + agent lint |
| Vector RAG | Huge corpora where grep and wikilinks fail |
| MEMORY.md in repo | Coding-agent session facts tied to one codebase |
| OKF bundles | Org-wide typed concept graphs across teams |
Below ~50K–100K tokens of dense wiki pages, Karpathy and implementers argue pure context beats naive RAG. Obsidian vaults often sit under that threshold per domain.
Related reading
explainx.ai guides
- Karpathy LLM Wiki: Agent Memory Pattern
- Google Open Knowledge Format (OKF)
- What Is MEMORY.md?
- Files.md: Local-First Notes
- Claude Code Loops Official Guide
- Build a Personal AI System
- OpenKnowledge / Inkeep local-first wiki
Primary sources
- Obsidian Help — Vaults
- Karpathy LLM Wiki gist
- chewa. viral thread (Jul 5, 2026) — with Community Notes context
Summary
The 1.1M-view "Anthropic engineer leaked Obsidian vault" post is not verified — Community Notes link the underlying article to a Porto science journal editor, not Anthropic. The video shows Obsidian Graph View, a static wikilink map, not a neural network running company decisions.
An Obsidian vault is a folder of markdown files you own, with an app for navigation and plugins. The productive takeaway is the self-writing vault: CLAUDE.md conventions, folder semantics, ingest/query/lint loops, and optional /schedule maintenance — patterns explainx.ai documents via Karpathy's LLM Wiki and OKF.
You can open Obsidian tonight, create a vault, add ten notes, link them, and point Claude Code at the folder. That is the real setup — no leaked brain required.
Practical test: after one week, ask Claude "What contradictions exist between my project notes?" If it cannot answer from 20-wiki/ and index.md, your ingest rules need tightening — not a bigger graph animation. The vault pays off when lint catches drift, not when graph view looks like a brain scan.
Community Notes status, Obsidian pricing, and referenced repo layouts reflect publicly available information as of July 7, 2026.
