My second brain is a folder of markdown files an AI keeps for me

The most useful system I run isn't an app. It's a Dropbox folder of markdown files that every AI session I start can read - and, more importantly, write.

I've kept notes for years the way everyone has: scattered, decaying, write-only. What changed is the reader. When your notes' primary audience is a language model that will sit down with them hundreds of times a year, the design constraints flip, and suddenly a knowledge base is worth maintaining - because it maintains itself.

Andrej Karpathy sketched this idea - an LLM-first knowledge base - and I built my life on it. Here's what mine looks like after daily use.

The golden rule: the AI writes, I read#

The wiki has one operating principle: the LLM writes the wiki. I read it. I curate sources and ask questions.

When a session figures something out - a vendor's contract terms, why we chose a platform, who a person is and why they matter - that finding lands in a note before the conversation ends. The default pattern for any non-trivial answer is: answer me, then file it. Karpathy calls this compounding answers. Good answers should leave residue instead of evaporating into chat history.

A year in, that residue is the asset: hundreds of topic notes, over a thousand person stubs, a full registry of my department's Hall of Fame, every active contract with structured metadata, and a chronological log of everything the system has ever done.

What LLM-first design actually means#

A few concrete choices that only make sense when your reader is a model:

Every note ends with a Context Block - one paste-ready paragraph that primes a cold AI session with the note's key facts. Human notes bury the summary. LLM notes lead with retrieval.

Strict frontmatter on everything. Status, dates, sources, people, tags from a controlled vocabulary. Machine-readable metadata means any session can query "every active contract" or "everyone connected to this project" without re-reading prose.

Atomic and heavily cross-linked. Many small notes with wiki-links beat a few big documents, because a model assembling context pulls exactly the notes it needs - and nothing else.

Uncertainty is captured, not buried. Every note has open questions. A knowledge base that only records what's known quietly lies about what isn't.

Single source of truth. Each fact lives in exactly one note; everything else - indexes, dashboards, queries - is a view. Duplicated facts drift, and a model can't tell which copy is right.

The ingest loop#

New material enters through a dump folder. I drop in whatever - a contract PDF, a meeting transcript, an email thread. The AI reads it, discusses it with me, updates every note it touches, appends to the log, and archives the source. The wiki gets richer; the pile of unprocessed stuff stays at zero-ish.

The same loop wraps meetings: prep doc before, notes filed after, action items synced. The system compounds precisely because nothing is allowed to end as "just a conversation."

Do you need this?#

You need three ingredients: files an AI can reach, a written schema the AI follows (mine is a single instructions file it reads at the start of every session), and the discipline of the golden rule - stop taking notes yourself and start making the AI file what it learns.

Plain markdown in a folder turns out to be the perfect substrate: no lock-in, human-readable, versionable, and every AI tool on earth speaks it. The stack is boring. The compounding is not.