A Company AI Brain
for Non-Techies

Zoom into one of those conversations and the rhythm is always the same. When I open Customer Magic – my outreach workstream – it reads its own files first: the docs that hold our rules, the running journal, the live dashboard. Then it pulls the truth from my CRM, inbox and calendar, and gets to work. It's not a task I kick off in the morning and close at night – it's a loop I run many times a day, and every time it finishes something, it writes back.

Every one of those conversations leaves the next one a current, honest picture – because the last session wrote it down. That's the whole trick, and it's worth saying plainly: the brain isn't a feature of the model. It's the documents these conversations load and update. Swap the model tomorrow and the brain stays. It's the same discipline I leaned on when I put our growth strategy in a codebase – the value lives in the writing, not the tool.

So what is that brain, exactly? Not a database – just a folder of plain files each conversation reads and writes. We write docs; the sharp ones harden into skills; a running journal holds the working memory; and dashboards turn all of it into something you can read at a glance. Four kinds, and you already understand every one.

Researchers dress these up as “working, episodic, semantic and procedural memory.” You don't need the jargon. You need four folders. And here's the quiet consensus most founders miss: the loud tools this year – Mem0, Letta, the vector and graph stores – are real, but they're specialist add-ons, not the default. The same month everyone's timeline filled with agent-memory courses and framework launches, the setup that keeps winning the head-to-heads on cost, editability and not-breaking is almost dull: plain files a session reads at startup, with the heavy machinery bolted on only once you truly outgrow reading your own notes. I'm not behind for running my business on documents. Documents are the frontier – the same way an everyday content skill beat a fancier pipeline for us.

The reason this compounds is that the brain edits itself. Here's a real example. For a while, my Customer Magic agent would finish an action – send a follow-up, say – but leave the matching task in the CRM sitting open. Small thing, but it added up: a pile of “open” tasks that were actually done. A reliability gap.

So we fixed it once, in the right place. I wrote a rule into the skill itself: finish an action, close its task, open the next one. Every session since enforces it. The brain didn't just remember a fact – it patched its own weak spot, permanently.

Now multiply that by every week. Each time we find a better way, it goes back into the playbook or the skill, dated, so the next session runs a little sharper than the last. I'm not re-explaining myself over and over. I'm reviewing edits to a brain that's teaching itself. That same brain is what let a single agent conversation research, build and ship a whole pitch deck – it already knew the company.

None of this is infrastructure. The unlock is two boring things: plain-text documents, and giving the agent broad access to the tools you already use. You steer and review; the agent does the reading, the writing and the grounding. This is AI founder mode for the rest of us – leverage without a lab.

It's not hands-off, and I won't pretend it is. Two things stay firmly with me. First, judgment and the money calls: the agent drafts the outreach and makes the small operational decisions, but nothing gets sent, and no real revenue call gets made, without me – the skills hard-stop and wait. Second, the one real trap: a brain made of documents is only as good as its consolidation habit. Let the docs go stale and the agent will confidently act on something that's no longer true. The fix isn't fancy – it's discipline: keep it current, and verify before you trust.

So if you've been sitting out the “agent memory” conversation because it sounded like an engineering problem, it isn't one. It's a writing habit with tools attached.

You're not managing a database. You're growing a brain that maintains itself – one document at a time, in language you can read, about the business only you understand. Start with one folder. Tell your agent to write things down. Give it access to one real tool. Then watch what it does before it shuts down for the day.

We trade a lot of these setups in the community – if that's your thing, join one of the upcoming Pirate Labs and you'll get pulled into the group.

Cheers,

Ben