Sancho
The second brain I work inside every day, built on the governance model I design for AI agents
Sancho is my work memory: one git repository of markdown files that any AI can read, navigate, and write into, with me as the only thing that makes anything permanent.
Markdown is the memory. Git is the history and the undo. Whichever model I'm using that day is the reasoning layer, and none of them are load-bearing. I've been working inside it every day since May 2026.
Concepts, decisions, specs, research, daily logs. One vault, so ideas can cross-link.
Claude, ChatGPT, Codex, Cursor. The rules live in markdown, never inside one vendor's app.
Daily, as my actual working memory rather than an archive I visit.
46 concepts, 18 decisions, 9 research bases. Counted July 25, 2026.
Three problems, one answer.
My memory is limited. I can't hold the full state of a governance framework, a shipped approvals product, an AI workforce model, and a portfolio in my head at the speed the work moves. AI memory is worse: every fresh session starts blank, so the first ten minutes go to the model rebuilding context out of Slack, Figma, and old chats. Every time.
And raw archives lie. I had 22,000 notes in Apple Notes, which sounds like knowledge and is actually a slower search engine. A pile of everything tells you nothing about what's true.
The name is the fix, stated. Sancho Panza is the squire who carries what the knight can't: the food, the facts, the memory of what actually happened. He's loyal but not a yes-man, he pushes back when he sees a problem, and he's grounded exactly where the boss isn't. The one thing he never does is grab the wheel. That's the contract, and any AI that opens the repo reads it first.
An archive is not a brain.
Underneath the contract, six decisions do the work. Each one was made because the obvious alternative would have quietly cost me control of what's true.
Markdown and git. Portable, permanent, readable by any tool and any human. Obsidian is an optional window, never a dependency. No app lock-in, because I bounce work between models on purpose and that only works if my memory isn't trapped in one vendor.
The model reads a compact index file at session start, then greps the repo on demand. No vector database. At this scale plain search wins, and every result is explainable.
An AI never invents a folder. Every file it writes is one of a closed set of types: concept, decision, spec, note, prompt, log, source. I say "save this" in plain language and the system classifies by meaning. If a file is none of those types it's probably junk, and the rule says don't write it.
Every note carries two labels. Status is the trust axis: seed for AI-written and unreviewed, hypothesis for something I hold but haven't validated, canonical for what I stand behind, deprecated for retired and never deleted. Horizon is the time axis: now, next, later, shipped. Nothing is ever canonical because an AI wrote it. It's canonical because I kept it.
A pre-write hook blocks any agent from editing the governance files or committing anything. Agent activity logs live in their own quarantined folder, stamped as machine records, never mixed with my thinking.
Every ceremony was run by hand first. A morning health check reports broken links and stale notes, silent on a clean day. On Mondays a ranked packet proposes connections, cleanups, and closeout candidates, each with its evidence, and I give one yes that applies exactly that.
Nothing here needs a terminal. I work in chat with whichever model fits the task and describe what I want in plain language. Six moves cover almost everything.
Files a note as a concept, decision, or spec with the right labels. Durable thinking gets distilled in; ephemeral gets a one-line pointer.
The AI interviews me until what's only in my head is on disk as raw material. This is the one I underestimated and now use most.
Turns a folder, transcript, or PDF into clean seed notes, with every quote checked back to its source mechanically rather than remembered.
The Monday packet: proposed links, syntheses, and closeouts, ranked, each carrying the evidence for why.
A read-only engine reporting missing owners, stale notes, and broken links. It checks and proposes. It never decides.
The hook that stops an agent editing its own rules. The one skill whose whole job is to say no.
Here's the part I didn't plan and then leaned into completely.
My day job is designing governance for AI agents. The framework I take to leadership says an agent proposes an action, the system classifies it, a gate decides, and execution is recorded. My second brain runs the identical shape at a scale of one. Same model, two sizes, which means every day I use it I'm also stress-testing the argument I make at work.
In practice it's undramatic. After a good conversation I ask for the artifact, not the transcript, because one clean concept note beats twenty raw chats. The AI files it. I read the diff, green for added and red for removed, and click commit on what deserves to survive. That click is the whole system. It's the moment something stops being a suggestion and becomes canon, and it's the only moment that can't be delegated.
When people ask whether the governance model is practical, that's my answer. I've been living inside it since May.
I don't just design governed AI collaboration. I run my life's work on it.
Two labels on every note means the whole repository is a map of what I trust and how far out it points. Plotted on those two axes, the shape of my own thinking is legible, including the parts that are uncomfortable.
Counted from the repository on July 25, 2026, and moving. The map below covers the mappable subset: concepts, decisions, and notes, with daily logs and READMEs left out. The notes in it are sample stand-ins, not my real corpus.
Pick a square above to list its notes, then click a note to read it and walk its links.
Because it's design work. Not screens: a system where a human and any number of AI models share memory without the human losing track of what's true. Every interesting problem in it was a design problem. What deserves to be remembered. How trust gets made legible. Where the boundaries go, and what an AI may never touch.
That's the same craft as my product work. It's just pointed at my own life, which is also the only way to find out whether the argument survives contact with daily use.
The enterprise version of this argument.
Delegated Authority FrameworkThis page is the story. The dossier is the record.
The full dossier carries what the page compresses: the architecture, the full governance model, the maintenance layer, the evolution history, and what the system deliberately is not. Written for the AI you'll paste this page into. Open to everyone.
www.arnoldp.com/dossiers/sancho.md
