Field note · The AI second brain

A second brain you don't have to build.

Tiago Forte named the second brain. Andrej Karpathy handed it to the AI. Garry Tan ran it across thousands of files. Between them they proved something real: an AI that maintains your knowledge for you works. And every single version still expects you to be a software engineer. Clone the repo, write the config, wire the API keys, schedule the agents, keep the vault synced and debug it when something goes wrong. Maxy gives you the same outcome with none of that work, on a graph that outperforms a folder of text files.

A short lineage

Four people turned a note-taking habit into an AI architecture.

In under a decade the "second brain" went from a productivity method to a pattern an agent maintains for you. Obsidian made local markdown its home throughout. The idea is proven; credit where it's due.

2017 →

Tiago Forte

Coined Building a Second Brain. Capture, organise, distil, express: your notes as an external mind.

2026 →

Andrej Karpathy

The "LLM Wiki": markdown maintained by the agent, keeping the index in context rather than retrieving it on demand.

2026 →

Garry Tan

Operationalised it. GBrain & Gstack run the wiki across thousands of files. "The agent runs while I sleep."

Now

The builders

Open kits (e.g. obsidian-second-brain) add self-rewriting pages, contradiction sweeps, nightly consolidation.

The catch

Brilliant, if you're a software engineer.

Read the setup guide for any of them and the friction is the same. The intelligence is real; the on-ramp is a developer's. Before the second brain does a thing for you, you do all of this:

What they ask of you

  • Clone a GitHub repo and install a command-line tool
  • Write and maintain a config file that disciplines the agent
  • Wire up API keys (Grok, Perplexity, Gemini, Whisper) and pay each per use
  • Set up the background and scheduled agents yourself
  • Keep the vault synced across devices, and debug it when a rewrite goes wrong

And it's still just you

  • One vault, one person; no shared team memory
  • No private-vs-public split; no access control
  • No undo, no audit trail, no export-on-request
  • Search is keyword and link-following, not meaning
  • Relationships are untyped: a link, never "who founded what"
They proved the idea. They also proved the price of admission: months of a developer's weekends. Maxy exists to remove that cost. The store you end up with is better than a folder of markdown files.

Maxy, axis one

The outcome, with none of the engineering.

There is nothing to clone, configure or debug. Maxy arrives installed and running on your own server: the graph, the agents, the ingestion engines and the channels are all assembled. It runs inside your flat Claude subscription, so there are no metered API charges for research or transcription. The second brain that took those builders months of setup is, for you, already on.

You skip

  • The repo, the CLI, the config file
  • The API keys and their per-use charges
  • The cron jobs, the sync, the debugging

You get

  • A running system on your own private server
  • Specialist agents orchestrated by one Claude
  • Your existing notes (Obsidian, Notion and more) imported on arrival

Maxy, axis two

And the store underneath is a graph, not a pile of files.

A markdown vault is a good single-user, local, read-mostly store. Each of those qualifiers breaks the moment the store is shared, searched constantly and held to account. A flat file needs a vector database, a search engine and a graph library added by hand to match what Maxy has built in.

What you actually needMarkdown vaultMaxy graph
Find by meaning, not keywordAdd an external vector DBVector + keyword search, fused in one query
Typed relationships (who founded, who attended)Wikilinks are untypedValidated edge types, queryable both ways
No duplicates, no contradictionsBy convention, by handEnforced at the moment of writing
Many people, one knowledge baseOne vault per personTenant isolation in the database itself
Private vs. customer-facing answersTwo separate vaultsOne store, scope-aware retrieval
Undo, audit, GDPR erase & exportDeletion is destructiveSoft-delete, grace period, full export
"What did we believe last March?"Frontmatter gymnasticsLiving summary beside an immutable timeline

The verdict

The idea was never the hard part. The build was.

Forte, Karpathy and Tan showed the AI second brain works. They also showed it costs a developer's months to stand one up and keep it alive. Maxy removes that cost. What you end up running is not a folder of text files; it is an ontological knowledge graph: shared, searchable by meaning, audited, and yours.

Keep an Obsidian vault you love? Bring it. Maxy ingests it on day one. Markdown is a fine place to think. A graph you did not have to build is a better place to store what you know.

See it in your stack

A second brain that's already on.

How Maxy works →