A second brain my AI agents actually read.
The Agentic Brain is a git-backed Markdown vault — the shared long-term memory for me and my agents. Obsidian for humans, GitHub as the source of truth, synced across a MacBook and three Mac minis, with real write-governance and a compiling knowledge wiki so agents don’t corrupt it.
Models change. Tools change. The knowledge shouldn't.
Every agent started cold.
My agents were trapped in their own runtime memory — no shared, durable, verified knowledge base to draw on.
Each model I wanted to use — the OpenClaw agents, Hermes, Claude Code, Claude Co-Work, Codex — re-derived, re-verified, and re-lost the same facts in every tool. Months of career proof, client SOPs, business strategy, and research were piling up with no system to turn raw material into linked, reusable knowledge.
And a governance problem was waiting to happen: the moment agents could write to the source of truth, one sloppy edit could quietly corrupt it — and I’d trust it anyway, because it looked clean. I wanted one canonical brain every model could read, that I could edit like notes, that survived any single tool going away, and that had rules to keep a writable brain from rotting.
Four layers on one idea: the folder is the asset.
The real asset is the plain-Markdown folder, not the app. From there the system has four layers, each with a clear job.
File over app. File over AI. If Obsidian, Claude, or Codex disappear tomorrow, the brain is still mine — still readable, still complete.
— Founding principle of the vaultObsidian for humans, Git for agents.
Obsidian gives me linked notes, backlinks, graph view, and a Web Clipper for intake. GitHub is the durable source of truth for agents — universal, inspectable, versioned, and not dependent on Obsidian being open. Every machine holds a local clone; agents pull before they read or write.
Numbered domains, one front door.
Nine numbered top-level folders so humans and agents always know where a thing lives and who owns it. A root START.md is the front door every agent reads first; me.md is the canonical durable profile that runtime files defer to instead of duplicating.
A writable brain needs contracts.
A protected core only the lead agent or I edit directly; declared folder ownership; and four contribution states — canonical, draft, proposal, archive. Nothing is treated as truth until I approve it. The operating rule for agents: brain first.
Raw intake that compiles into knowledge.
An immutable raw/ intake folder, a fingerprinted source registry, and a compiled wiki/ layer where an agent extracts entities, concepts, syntheses, and open questions into interlinked pages. It compiles knowledge instead of just retrieving files, so the brain compounds over time.
Sync topology.
GitHub is the source of truth at the center. The MacBook Pro bridges Obsidian to Git in auto mode; the three agent minis run as pull-only clones so in-progress agent work is never auto-committed. An installed launchd job keeps all four in sync on a five-minute cadence.
Obsidian edits land on the MacBook Pro bridge, which commits, rebases, and pushes to GitHub — the one canonical source of truth (holisticseo/agentic-brain, main). The three Mac minis pull that repo in pull-only mode so agent work-in-progress is never auto-committed. An installed launchd job runs every five minutes on all four machines. A Gitleaks scan gates every sync, and conflicts stop for manual review rather than being silently resolved.
Vault & Format
Source & Interface
Sync Automation
Safety & Secrets
Governance
LLM Wiki Layer
One brain, replicated.
Structural counts verified from the live vault on the MacBook clone and the build thread.
The first LLM Wiki ingest compiled 24 registered sources into 9 entities, 11 concepts, 3 syntheses, and 4 open-question notes — roughly 51 interlinked pages, lint-passing for orphans and broken links. Every phase closed with a written QC record inside the vault.
A writable brain needs contracts.
A shared, agent-writable brain without rules becomes an unreviewed dump. So on top of the files sits a governance layer: a protected core, declared ownership, and four explicit states that gate what counts as truth.
The root instruction files — START.md, me.md, vault-map.md, skill-map.md, CLAUDE.md, the history-epoch marker — and all of 00-system/ are edited directly only by the lead agent (MC) or by me. Every other agent proposes changes. Each numbered domain has a declared owner and a “Mike-owned, agent-assisted” rule set.
Canonical
Approved truth. Nothing reaches this state until I promote it. Agents read it as authoritative.
Draft
In progress. Strategy, offers, and pricing stay here until reviewed — never treated as fact.
Proposal
Agent-authored change awaiting my sign-off. Client-facing claims and positioning live here first.
Archive
Superseded or parked material kept for traceability — out of the way, still versioned.
Brain first: check the vault before searching the outside web or reinventing context.
— Agent operating doctrineRaw intake that compiles into knowledge.
Adapted from Andrej Karpathy’s LLM Wiki pattern: three tiers that turn a raw drop into linked, reusable knowledge instead of a dead pile of files — so the brain compounds over time.
Raw drop
Anything — Markdown, PDFs, images, transcripts, web clips — lands in an immutable raw/ folder. Never edited.
Registered source
A source registry fingerprints each source with a SHA-256 hash via wiki-register-source.sh.
Compiled
An agent extracts entities, concepts, syntheses, and open questions into a wiki/ graph.
Linked wiki page
Interlinked Markdown, health-checked by wiki-lint.sh for orphans, broken links, and low-confidence claims.
Neutral worked example: “How Search Engines Work” — raw source through to a compiled, lint-passing wiki page.
What exists — and an honest status.
Two separate ledgers. The left is what is built, running, and verified in the live vault. The right is the honest status — kept deliberately distinct so proven facts never blur into roadmap.
A live, governed, git-backed second brain in daily use.
- One canonical GitHub source of truth, editable in Obsidian, read by all five agent runtimes.
- Kept in sync across four machines by an installed launchd automation — auto on the bridge, pull-only on the minis.
- Every sync is secret-scanned with Gitleaks; conflicts stop for review instead of corrupting truth.
- Governance assigns ownership and gates canonical promotion behind my approval.
- Populated and in use: career profile, a source-backed claim library, client folders, business context.
- A working LLM Wiki took its first source from a raw drop through to compiled, linked, lint-passing pages.
Young, actively-iterated infrastructure — not a mature product.
- Started June 2026: Phase 1 scaffold → GitHub sync → LLM Wiki → automated multi-machine sync → a supervised history rewrite.
- The LLM Wiki is still supervised manual ingest (Phase 1): I drop files, then prompt an agent to register and compile them.
- Automatic watching, extraction, and auto-compiling are planned later phases, not shipped.
- Garry Tan’s GBrain was evaluated and deliberately deferred — the file-native version ships first and stays canonical.
Framed correctly: living infrastructure I designed and operate, under active buildout — not a fully-autonomous, self-compiling, battle-tested platform.
From scaffold to running fleet.
MC scaffolded the numbered folders and seeded me.md and the vault map. Claude pressure-tested it — approved after fixes: misfiled notes moved, an unplanned folder removed, secret-ignore rules hardened.
The private repo became the source of truth; every machine took a local clone. Relative links pass, no stray wikilinks, main matching the clones.
First real source taken from a raw drop through registration and compilation to linked, lint-passing wiki pages.
A hardened brain-sync.sh plus a launchd installer — com.agenticbrain.sync — on a 5-minute cadence: auto on the bridge, pull-only on the minis.
A history-epoch marker guards against replaying purged commits after an intentional rewrite — sync refuses to rebase across a mismatch.
I owned the architecture and the hard calls.
I owned the vision and the hard decisions: file-over-app, GitHub-as-truth over Obsidian Sync for agents, the numbered-domain structure, the protected core, and the canonical/draft/proposal/archive discipline.
I made the call to adopt the Karpathy LLM Wiki now while deferring GBrain. I directed MC — my primary OpenClaw agent and the declared owner of 00-system — to build the scaffold, seed the profile and vault map, stand up the sync scripts and launchd automation, and run the wiki ingests, then reviewed and approved as it went, with Claude pressure-testing before sign-off. I set the guardrails: business strategy, client claims, and positioning stay draft until I promote them; secrets and unsanitized exports never enter git. I am the owner and final decision-maker; the agents are the builders I direct.