Most of us run two parallel worlds. An Obsidian vault with hundreds of notes nobody rereads, and AI agents in the terminal where every session starts from zero, as if the past never happened. The obsidian-second-brain repository glues those worlds together: 4.7 thousand stars, MIT license, authored by Eugeniu Ghelbur, an AI automation engineer. The README states its formula plainly: “an evolution of Karpathy’s LLM Wiki pattern: a vault that rewrites itself”.
This post is a review of the repo, and at the same time an explanation of why we read it with constant recognition: our Automata CRM runs on the same pattern, only for business entities instead of notes.
Karpathy’s LLM Wiki pattern
Andrei Karpathy described a simple, powerful idea: make your knowledge base a wiki maintained by an LLM. The model adds new pages and links them, finds contradictions and flags them for manual resolution, and surfaces patterns when you ask. The human owns the knowledge; the model is the workforce that maintains it.
The pattern’s weak spot is that it’s passive. The wiki grows when you ask. Contradictions sit until you resolve them. Obsidian Second Brain pushes the idea into its active form.
What’s inside: four layers, 47 commands
The repo gives a CLI agent (Claude Code, Codex, Gemini CLI and four more) a set of commands and background agents that operate on your vault.
Operations layer, 30 commands. Saving and ingesting material: URLs, PDFs, audio, screenshots become normalized notes. /obsidian-reconcile resolves contradictions, /obsidian-health lints the vault, /obsidian-architect documents a codebase into notes.
Thinking tools, 9 commands. The most interesting part. /obsidian-challenge makes the vault argue with you based on your own history. /obsidian-emerge hunts for unnamed patterns, /obsidian-connect builds bridges between knowledge domains.
Context layer. /obsidian-world loads identity and state with token budgets, from a light L0 to a full L3: the agent gets exactly as much context as the task fits.
Research toolkit, 7 commands. Search, digests, YouTube and podcast notes. The key word is vault-first: /research-deep scans your notes first and only researches the gaps in the outside world.
Plus four scheduled agents: morning at 8:00, night at 22:00, a Friday review and a Sunday health check. The vault lives without you.
Engineering discipline: OKM and bi-temporality
Two things separate this project from “let’s plug GPT into notes”.
Open Knowledge Metabolism. Every fact must be timeless, dated, or a pointer. Slow-changing knowledge lives in the note. Fast-changing facts point to a live source with an “as of” date. There’s a spec and a linter.
Bi-temporal facts. The vault remembers not only when a fact was true but when the vault learned it. It sounds academic until the question “so when did this stop being true” comes up, and it always comes up.
And the principle to start with: the AI-first format. Notes are written for an agent to read, not a human: frontmatter, recency markers, wikilinks, verbatim sources. It’s an inversion: for decades we adapted machines to documents, here documents adapt to the machine.
The engineering deserves respect on its own: 763 tests in CI, eight builds for different platforms, and bounded recall through a hook, at most four notes of roughly 900 characters per prompt, so the “memory” doesn’t eat the whole context window.
”Your vault outlives any CLI”
The most practical line in the README. Agents change every six months: one today, another tomorrow, a third the day after. Knowledge in markdown belongs to none of them. You switch CLIs, the vault stays, the new agent simply reads the same files. A rare case where the right architecture is also the cheapest one.
We built Automata CRM on this pattern
We kept reading with deja vu, because our Automata CRM is a production implementation of the same idea, on business entities instead of notes.
Everything below is not a roadmap, it’s how the system works today:
- The vault is the source of truth. Clients, deals, tasks, invoices, infrastructure and correspondence live in an Obsidian vault: markdown with frontmatter. A task tagged urgent and important lands in the right quadrant of an Eisenhower matrix on the dashboard, automatically.
- Agents read and write directly. AI agents work on the same files: a board of director roles (a CEO holds the weekly plan, a CFO holds cash and price floors, a commercial director holds receivables), a marketer holds promotion data. Decisions land in a shared journal, and a human approves what matters.
- Data stays in place, panels read it. Dashboards don’t replace the vault, they read from it: the promotion panel compiles semantics, content coverage, five social networks and traffic into one JSON file that both the agent and the owner can see.
- Data discipline. Sensitive client fields are encrypted in a separate vault that locks itself, and the company is a registered personal-data operator. For a business version of the pattern that’s not an option, it’s the survival condition.
How the business version differs
The difference between a “second brain” and a “CRM on a second brain” is the cost of an error. In a personal knowledge base, a wrong note costs a reread. In business, an agent that records a payment wrong or leaks personal data costs money and reputation. So our version triples down on what the personal version treats as optional: human approval of critical actions, encrypted personal data, and a decision log that always shows who did what and why.
The takeaway is simple. The “LLM plus wiki” pattern has traveled from a tweet to systems with hundreds of tests, and to business tools companies actually run on, all within a couple of years. If you already have an Obsidian vault full of knowledge, obsidian-second-brain is the shortest path to giving your agents persistent memory. If you’re a business owner thinking about a CRM you won’t have to buy back from a vendor along with your data, it’s the same pattern, and it works.
Phone: +7 (906) 311-77-69, email: hello@automata.sale, Telegram: @automatasale, site: automata.sale
Evgeny Uryadov, sole proprietor (INN 645112058391, OGRNIP 312645301900058). Working across Russia, Belarus and Kazakhstan.