Obsidian as working memory.

The full guide. PARA vs Zettelkasten vs Johnny Decimal vs five others — compared honestly with operator-grade verdicts. Vlad's project-as-entity reframe (the only method built for portfolio operators + AI agents). The neuron logic. The import workflow that turns Apple Notes + email into a firing brain. A working vault you can clone tonight.

Companion to Ch 4 — Obsidian as Working Memory. Built from three parallel AI research passes (methods comparison, Obsidian mechanics, Vlad's specific pattern), synthesized in his voice. Read in any order; everything stands alone.

Why Obsidian wins for AI-native operators

Four reasons, in order of how much they matter when AI agents are first-class readers.

Local-first. Your vault is a folder of .md files on your machine. Not in someone else's basement. You can ls it, grep it, rsync it to a backup drive. If Obsidian the company disappears tomorrow, the files still open in any text editor for the next 40 years. Notion can't say that. Reflect can't say that. Mem definitely can't.

Markdown native. Every model — Claude, GPT, Gemini, the open-weights you'll be running in 2027 — reads markdown natively. Zero parsing layer between your notes and an LLM. Point a filesystem MCP at the folder and the agent just reads. Notion requires an API token, scoped permissions, page-by-page sync, and rate limits that bite at the worst time. Obsidian gives the bytes directly.

Bidirectional links. Mention [[Mentee A]] in a session note and the link is alive in both directions. The backlink panel on Mentee A shows every note that references them. That's the neuron — one node activates the network. Notion has page links but they're not symmetric. Roam had this and got it right, then ran out of runway.

Plugin ecosystem. ~1,500 community plugins, open source, two-click install. When you need a thing, someone's built it. When nobody's built it, Templater lets you script in plain JavaScript. You don't wait for a roadmap.

Free. Optional sync at $8/mo. iCloud or Dropbox for $0. Mobile apps are decent — phone is for capture, desktop is for synthesis.

Eight methods compared, honestly

Before structure: ask which axis you're committing to. LATCH (Richard Saul Wurman's claim from Information Anxiety) says there are only five ways to organize anything — Location, Alphabet, Time, Category, Hierarchy. Every folder structure you've ever built is a mix of these axes, usually badly. If your vault feels wrong, it's almost always because you mixed axes at the same level. LATCH is theory, not method — read it once, internalize, don't try to "implement" it.

PARA — Projects / Areas / Resources / Archive

Tiago Forte's framework. Four buckets, every note lives in exactly one. Projects have outcomes and deadlines. Areas have standards but no end date (a company you own, your health). Resources are reference material on topics. Archive is dead projects and outdated reference. The logic: file by how actionable, not what topic. When a project ends, the doc moves to archive or gets promoted to a resource.

Where it shines: solo founders with 3-7 active projects. Consultants. Writers shipping one book. PARA is the best method for someone whose job is "produce deliverables."

Where it breaks: five companies, twelve quarterly initiatives, two mentees, a newsletter, an investor pipeline does not fit "projects vs areas." Everything you touch is both — a company is an Area (ongoing standard) AND a stream of Projects (Q2 hire, fundraise, ship product). Forte says: nest projects inside areas. But that defeats the premise. And relationships, deals, mentees, songs — they're entities with state, and PARA has no slot.

Zettelkasten

Niklas Luhmann's method, 90,000 index cards in a wooden cabinet. Digital version: atomic notes, unique IDs, mandatory linking, three intake stages (fleeting → literature → permanent). Value: stop organizing by folder, start organizing by connection. Every note is a node, every link an edge, structure emerges from the graph.

Breaks below 500 notes (graph too sparse for serendipity), breaks above 5,000 (graph becomes a hairball). Breaks for operators because Zettelkasten has no concept of state. "Ship v2" is not a Zettelkasten note. "Meeting Chris next Tuesday" is not a Zettelkasten note. It's a method for thinking, not for running things.

Verdict: keep a small Zettelkasten subfolder for ideas you'll cite ten times. Don't run your whole vault on it.

Johnny Decimal

Numeric prefixes — 10-19 areas, 11.01 / 11.02 / 11.03 items. Filing-cabinet brain. Find anything by reciting the path: "Q2 board deck — 30 finance, 32 board, 32.04."

Shines for static reference (contracts, financial docs, legal). Breaks the moment you have an 11th category in an area. Doesn't handle entities-with-lifecycles. AI agents don't care about your numbers.

Verdict: great for the 10% of your vault that's static reference. Terrible as a whole-vault method.

GTD-based vault

David Allen's framework adapted to Obsidian. Top level: Inbox / Next Actions / Waiting / Someday / Projects / Reference. Genius for tasks. Disastrous for knowledge — forces every note into the to-do shape, your vault becomes a guilt machine.

Verdict: run GTD in a tasks tool (Linear, Todoist, Things). Let your vault be a vault.

Bullet Journal applied to digital

Ryder Carroll's analog method. The daily-note rhythm and the weekly migration ritual survive the digital conversion. The bullet symbols and the manual index don't. The whole point of BuJo is the friction — rewriting a task tomorrow forces "is this still worth doing?" Digital has no friction.

Verdict: steal the daily-note + weekly-review rituals. Skip the symbols.

Project-as-entity (Vlad's twist) — the method that actually works at portfolio scale

Forte defines a project as "an active outcome with a deadline." Too narrow for anyone running more than one thing. The right primitive for portfolio operators is broader:

Under that definition:

• A company is a project. Belkins has state (MRR, headcount, runway), a lifecycle (founded → growth → exit), decisions, history. It's not just an "area" — it's an entity you actively shape.
• A person is a project. A mentee has state (mentoring phase, current homework, last session date), a lifecycle (intake → growth → graduation), decisions, history. Treating them as a "resource" loses the entity-ness.
• A deal is a project. Stages, decisions, a clock, a history.
• A quarterly initiative is a project — classic Forte definition.
• A song, a book, a talk, a furniture build — projects, even without formal deadlines.
• A relationship is a project. Controversial framing but accurate — state, decisions, history, deserves the same structural respect as a company.

Top-level structure becomes entity-typed, not action-typed:

Vault/
├── 00-Inbox/        (capture zone)
├── 01-Daily/        (time-indexed)
├── 02-Projects/
│   └── Active/      (entities with state)
├── 03-People/       (one file per recurring human)
├── 04-Companies/    (portfolio entities)
├── 05-Newsletter/   (publishing pipeline)
├── 06-References/   (evergreen, non-entity)
└── 99-Templates/

Why this works for AI agents: when an agent reads your vault, the most useful question is "what's the current state of X?" Entity-typed folders make that a one-path lookup. PARA forces the agent to guess whether X is currently a Project or an Area, and the answer depends on time. Project-as-entity is a stable mapping. An agent reading 03-People/Mentee A/ finds the hub note, the action tracker, the session log, the patterns file — and can answer any question without crawling six folders.

Why it scales past PARA's ceiling: PARA breaks at ~15 active projects because the Projects folder becomes a list you can't visually scan. Entity-typed folders break at ~50 entities per type, which for most operators is years away. When an entity goes dormant (a deal dies, a relationship fades), it migrates to that entity's archive subfolder, not to a global archive that mixes everything.

The catch: requires committing to entity types up front. Add a new top-level type and you're moving folders. Pick 6-8 types and stick with them.

The 15-minute install

1. Download Obsidian — obsidian.md. Free. Mac, Windows, Linux. ~80MB.
2. Create a vault folder. I do ~/Desktop/Obsidian/Vlad-Brain/. On the desktop because you want it visible, not in a path with a space (some scripts choke on spaces). Pick a brain name. Mine is "Vlad-Brain." Sounds dumb. Works.
3. Point Obsidian at the folder. Obsidian → Open folder as vault → select the folder.
4. Sync. Three options:
   • iCloud — free, encrypted at rest, Mac+iPad only. Vault goes inside ~/Library/Mobile Documents/iCloud~md~obsidian/Documents/.
   • Dropbox — free, cross-platform, NOT end-to-end encrypted. Dealbreaker if your vault has client PII.
   • Obsidian Sync ($8/mo) — end-to-end encrypted, native, instant. If your vault has anything sensitive, this is the path.
5. Mobile. Install on iOS/Android, point at the same vault folder, you're synced.

First 5 minutes of use: create a note 00-home.md. Type a sentence. Cmd+E to toggle preview. Close Obsidian, open the same file in TextEdit — same content. Nothing is locked inside the app.

The 8 plugins, the 3 to avoid

Settings → Community plugins → enable → Browse → install each.

Three to avoid (the trap plugins)

• Kanban — looks cool, becomes tech debt. use Linear or Notion.
• Mind Maps — all implementations clunky — use Whimsical or Excalidraw full canvas.
• Sliding Panes — visual chaos past 3 panes. built-in tabs win.

Install plugins that disappear into the workflow. Avoid plugins that ARE the workflow.

Neuron logic — how the graph becomes a brain

Most people see the graph view as a pretty picture. That's the screenshot version. The graph as a diagnostic tool is what makes it worth the discipline.

Look for four patterns:

• Dense central cluster = your daily-driver themes. The things you think about every week. Good. This is the brain firing.
• Hub nodes = key entities with many threads (a company, a mentee, a long-running project). Click one, the whole context activates. Good — your themes are deepening.
• Radial outer ring of orphan dots = one-shot captures that never got linked back. Warning sign. Either link them or archive them.
• Floating clusters disconnected from the center = entire themes you've stopped engaging with. Decide: revive or archive.

The neuron-firing pattern: click [[Mentee A]] → activate all session notes, action items, patterns, strategic map. One click reveals the entire relationship context. Compare to Notion's database row: same data, no firing pattern. The graph IS the firing pattern.

This is why AI integration works. An agent reading a hub node gets the whole context in one read. An agent reading a Notion database has to do N queries.

Every two weeks, scan the graph for new orphans (red flag), new hubs (good — themes deepening), and dense clusters that should be split into multiple entities.

Import workflow — from Apple Notes + email to a firing brain

I did this. Uploaded every Apple Note + key emails into 00-Inbox, mapped from there. Took ~4 weeks to feel real. Here's the playbook:

Step 1 — Export Apple Notes

Notes app → File → Export, or third-party tools like Exporter (Mac App Store) or the notes-to-markdown CLI. Result: a folder of .md files, one per note. Lossy formatting on tables, embedded sketches, voice memos — accept the loss, archive the lossy bits, move on.

Step 2 — Export key emails

Use Mimestream / Apple Mail's "Save As" or Gmail Takeout. Don't import everything. Import: client emails, mentee threads, investor updates, key decisions. Skip: receipts, marketing, newsletters. The 80/20: probably 200-500 emails matter from your last 3 years. Save each thread as a single .md with subject + participants + date in frontmatter.

Step 3 — Drop everything into 00-Inbox/

Don't try to organize during import. Inbox is the dump zone by design. After Step 2 you might have 2,000 notes — that's the raw matter stage.

Step 4 — Bulk-link the obvious entities

For each note, identify the 1-3 main entities mentioned (company names, people names, projects). Use Templater + a regex pass — or Cowork with filesystem MCP — to auto-add [[Entity]] links where names appear. This is where the graph starts firing. Even rough links create the network.

Step 5 — Promote Inbox → permanent home

For each Inbox item, decide:

• Daily note? → 01-Daily/YYYY-MM-DD.md
• Project artifact? → 02-Projects/Active/<Project>/
• About a person? → merge into 03-People/<Name>.md
• Reference? → 06-References/
• Trash? → delete

Most notes find a home in 1-2 minutes. Some Inbox items spawn new projects ("oh, this whole Apple Note IS the Mentee A project").

Step 6 — Week-2 audit

Run the graph view. Dense clusters (good), orphan rings (need linking), disconnected clusters (themes you forgot about). Pick the top 3 hub candidates → turn them into proper Project notes with metadata frontmatter (status, last_updated, owner, open_threads).

Step 7 — Daily capture from here forward

00-Inbox stays the dump zone. Daily note fires from Calendar plugin. Weekly 20-min review empties Inbox, archives stale projects.

Wiring Claude Code and Cowork to the vault

The filesystem MCP server points at the vault folder. The agent reads your vault as text.

{
  "mcpServers": {
    "vault": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-filesystem", "/Users/<you>/Desktop/Obsidian/<Brain>"]
    }
  }
}

Drop that in ~/.mcp.json (Claude Code) or wire it via Cowork's Connectors panel. Restart, run /mcp, ask: "Read CLAUDE.md and tell me what's on my plate today." First useful answer in ~10 seconds.

Read-only by default. The agent should be able to read your vault but not arbitrarily write to it. Wire write access only when a skill genuinely needs it (the morning-briefing canvas writes to a specific output file; the mentee fan-out writes to a specific tracker — both narrow scopes).

The CLAUDE.md at the vault root is read on every session. Keep it under 100 lines. Ch 4 covers this; the working starter is in the clonable repo (link at top).

The loop that keeps it alive

A vault that doesn't get written to is a vault that rots. Without maintenance your "second brain" becomes a graveyard.

• Daily — 5 min. Write today's daily note. Anything that crossed your mind that isn't tied to a project goes in 00-Inbox. Don't organize. Don't link. Just capture.
• Weekly — 20 min (Fri afternoon or Sun night). Empty 00-Inbox into permanent homes. Archive stale projects. Update CLAUDE.md if focus shifted.
• Biweekly — graph scan. Look for orphan rings + disconnected clusters + hubs growing too dense (split into sub-notes).
• Monthly — consolidate. Merge duplicates. Prune dead notes. Fix broken links.
• Quarterly — vault audit. Does the structure still match how you work?

Year 1 feels like overhead. Year 2 you have an orchard. Every AI instance you spawn benefits from years of accumulated decisions, conversations, patterns, mistakes.

Five vault failure modes (skip if you've already lived them)

1. Starting with 12 top-level folders. If your root is wider than 8, you don't have a structure — you have a pile.
2. Using tags AND folders for the same hierarchy. Pick one axis per question. Don't tag #client if you fold by client.
3. Making Obsidian your task manager. It isn't one. Tasks live in a tool that pings you. Notes live in a tool that doesn't.
4. Renaming a person note when their title changes. Breaks all backlinks. Update frontmatter, leave the filename.
5. Refactoring more than twice a year. Every refactor breaks links, breaks muscle memory, breaks the agents reading the vault. Pick a structure, live with its weaknesses.

Which method, by shape

• Solo founder shipping one product → PARA. Ship by Friday, refactor in six months.
• Multi-company CEO (3+ portfolio) → project-as-entity with PARA nested inside each entity. PARA alone collapses by month 8.
• Paid mentor → each mentee gets their own entity folder under 03-People/ with hub note + action tracker + session log + patterns file.
• Newsletter operator → PARA for active issues + a small Zettelkasten subfolder for evergreen claims you'll cite repeatedly.
• Researcher / essayist → Zettelkasten as primary, PARA as a thin top layer for active deliverables.
• Freelancer with rotating clients → PARA wins. Clients are projects, skill areas are areas.
• Content creator (video / podcast / social) → project-as-entity with Episodes/ or Videos/ as a type.

The methods change. The discipline doesn't. Pick a structure you can defend to your future self in October. By April, half the tools above will have shifted tiers. Your folder tree won't.