Skip to content
Own build · my context layer

Knowledge an agent can act on.

I hand real work to AI agents: meeting prep, follow-ups, a brief every morning. An agent can only do that work if it knows what I know. What was decided on a call. Who someone is and how I know them.

It started small. In fall 2025, when Cadre moved to Claude, I used Projects and skills to turn messy transcripts and working files into structured records. It worked, one project at a time. Each project only knew its own material. Claude couldn't edit an old record, so every correction became a new file saying the old one had changed. And I had little control over how it searched.

So in April 2026 I had Claude write one wiki from all my meetings, articles and notes, following Andrej Karpathy's LLM wiki pattern. It grew to thousands of pages. I had marked 1 of 3,818 as checked. An agent can't tell which pages are wrong. It acts on whatever it reads.

I rebuilt it as Brain. Brain stores single claims instead of pages. Each claim is tied to the exact words it came from, and code checks it before an agent can use it. The goal is knowledge reliable enough that work built on it needs almost no correction from me.

The rest of this page covers how Brain works, how I tested it, what runs without me and where it falls short. It ends with Canon, the version I designed for a whole company.

Brain on August 20, 2026

sources
0
claims, each tied to a quote
0
wrong, and 3 unclear, in 282 claims used in real meeting briefs
0

Numbers only. The content stays private.

How it works

Every new meeting, article or note goes through four steps overnight.

  1. 1 · Capture

    The source is saved as is.

  2. 2 · Extract

    A model pulls out single claims, each with the exact words it came from.

  3. 3 · Check

    Code confirms the quote is in the source and long enough to carry the claim. Then a second model, the judge, rules whether the quote supports it.

  4. 4 · Serve

    Agents get supported claims only. Stale ones are held back, and the answer says so.

At the last count, 90% of claims were supported (47,791 of 52,982). The rest are kept and labeled, never served as fact. Supported means the source said it. If the source is wrong, the claim is wrong too.

Every write goes through one guarded function, and OpenAI's Codex reviews each milestone as an outside auditor.

How I tested it

The judge is a model grading another model's work. So I tested the judge against my own rulings.

The blind test

I took 20 real claims and made 70 fakes by corrupting real ones in six ways, listed in the table. Then I labeled them blind. I stopped at 49 of the 90, once every kind of fake had been sampled. Of the 29 fakes I ruled on, I caught 26.

Then the judge took the same test. It accepted all 20 real claims. Of the 70 fakes it caught 42, passed 27 as true and was unsure on one. It caught every changed number and every flipped meaning. It was worst on thin quotes, 2 of 12.

The thin-quote misses came from the setup. The judge saw 600 characters around each quote, text the agent never gets. Shown the quote alone, it caught 7 of 12.

I had hoped the judge would catch what I miss. It missed the same three fakes I did. So code handles what can be checked mechanically, and a person makes the judgment calls.

The last check was on real work. Four meeting briefs cited 282 claims. The quote alone settled 186. AI reviewers read the other 96 against the full source, to a standard I set. None were wrong. Three were unclear, like a tentative date stated as firm.

Fakes the judge caught, by kind
Changed number12/12
Flipped meaning or swapped name11/11
Wrong speaker7/12
Stretched paraphrase6/11
Overclaim4/12
Quote too thin2/12
All fakes42/70

Four smaller tests

Against the wiki
45 sources had both a wiki page and a Brain rebuild. An AI scorer compared each pair. 98.4% of Brain's statements traced cleanly to a quote, against 82.1% of the wiki's. Brain matched or beat the wiki on 44 of the 45. The other was mixed.
Against my own labels
Before the blind test I labeled 37 hard claims and compared the check step with my labels. It was rejecting true claims. The cause was a code rule on quote length. Loosening it doubled the share of true claims accepted, from 0.34 to 0.69.
Extended thinking, on and off
53 labeled claims, judged both ways. Thinking cost 2.3× as much. It accepted one more true claim and approved one false one. It stays off.
Against live tools
31 tasks, done from Brain, from live tools reading my email, Slack and documents, and from both. I wrote my predictions first, and 29 held. Brain was better for decisions, people and history. Live tools were better for anything current or inside a document. One task planted a deal that never happened. Both refused to invent it.

What runs without me

Checked claims are what let me stop reviewing every output. If a mistake would stay contained, the work runs on its own. If it could get out, the work waits for me.

Contained and reversible

Runs on its own

  • Processing every new meeting, article and note overnight
  • Checking each new claim against its source
  • Rechecking claims that have gone stale
  • The morning brief, which ships even when the news is bad
Escapes or can't be undone

Waits for me

  • Anything sent outside: email, Slack, posts
  • Changes to the system's own rules
  • Anything I verified by hand. A machine never deletes it.
  • Judgment calls the checks can't settle

Each kind of work starts supervised and moves across on its record. The number I watch is factual errors in finished work that nobody caught. The target is zero.

Where it falls short

Where things stood in August 2026.

  • The judge passed 27 of the 70 fakes as true and was unsure on one. The fix, limiting what the judge sees to the quote, was still open.
  • Brain held less than the wiki. An AI scorer put it at about 54% of the facts in each source, against 69% for the wiki, because anything without a quote is left out. Summaries come next, and they have to pass the same checks.
  • The samples are small. Dozens of labels and single runs, not thousands.
  • It hadn't replaced the wiki. Brain ran nightly, and I still used the wiki every day.

Canon, the company version

Brain can say whether a source said something. It can't say whether that is still true, who decided it, or who is allowed to see it. A company needs all of that. So at Cadre I designed Canon, one shared record that people and agents both work from. It is a design. None of it was built.

In the design, every answer carries a status that tells an agent how far it can take it. Pick a question to see each one.

Ask it something

Invented examples, in the shape Canon specifies.

Readyas of Oct 2

HubSpot. They finished moving off their old CRM in May.

The official tool record and the people doing the work agree, and both are recent.

Safe to use and act on within your permissions.

Evidence

  • System recordOfficial record

    “HubSpot · approved for sales · May 12”

  • MeetingSupports it

    “We finished moving everyone over to HubSpot last month.”

An agent may

  • Use it in meeting prep
  • Draft an integration plan for review

It may not

  • Change anything in their CRM. Reading isn't permission to act.
Ready
Use it and act on it, within your own permissions.
Limited
Only the narrower use it names, like a draft.
Review needed
A named person decides before anyone relies on it.
Blocked
Nothing comes back, not even that it exists.

The status guides the agent. The tools that act check again on their own and never trust the label. Canon also checks permission every time an answer is served, and it tracks who decides something separately from who knows about it.

What I took from it: decide who may see what, and who can make something official, before the first real data goes in. It can't be added later. When I left Cadre, Canon was a confirmed design: 22 spec documents, no code, no data.

Let's compare notes.

If you're building something like this, I'd like to hear about it.