Issue #17 · Ops & Om · Health of Business series · Episode 2
or listen on Spotify · Apple Podcasts
I want to tell you about the day my AI agent told me I was wrong, explained why — and then admitted it had been wrong first.
I use an agent to run paid ads for clients. The setup was the easy part. I connected the accounts, briefed it like the best media buyer I could hire — and then spent months going in circles. Not because it was bad at the job. Because every six weeks I'd start a new chat, and it would forget everything we'd learned.
You know the moment. The chat runs out of room, so you open a new one. And even inside the same project folder, the new one has a kind of amnesia. Every campaign I'd re-explain, re-correct, watch it re-derive the same reasoning to the same conclusions.
I wasn't undisciplined and I wasn't using the wrong tool. I was the memory.
Most people who quietly gave up on AI hit exactly this and decided the tool didn't work. It worked fine. It just couldn't remember — and nobody told them that remembering was now their job. So I built it a memory. Here's the short version.
Here's what we'll cover in this issue:
The three-layer memory that fixed the amnesia
The day the memory was wrong (two errors)
The job you can't automate away
The Three-Layer Memory That Fixed the Amnesia
The whole thing is three layers, and they do different work.
The chat is where the work happens — and where it dies. That's not a flaw, it's a container with a size. Everything you want to survive has to leave before the chat does.
The project instructions hold how it behaves — the persona, the standing rules, the guardrails. This survives the chat, and it should be read before every answer.
The history is a doc (or a Google Sheet) in Drive. Every campaign — what we ran, what we spent, what we expected, what happened, what it changed our minds about — gets logged. And one standing instruction ties it together: before you do anything, go read the history and pull what's useful for this campaign.
That third layer changed everything. Process time dropped to a fraction. The agent stopped re-deriving from zero and started every campaign already knowing what we'd learned.
The memory accretes. It compounds instead of drifting.
Try this prompt:
Create a running "campaign history" doc for this project. After each
campaign, append an entry with: what we ran, what we spent, what we
expected, what actually happened, and what it changed our minds about.
Then add this standing rule to your instructions: before proposing
anything new, read the full history doc and pull every lesson relevant
to this campaign.The Day the Memory Was Wrong
In June we ran a hero offer — buy two, get one free. $50/day, cold broad audience, dedicated landing page. It plateaued at 2.2x ROAS, CPA $78, frequency 6.77, reach stuck near 13,000. Six days left. The agent's advice: ride it out, don't refresh, don't touch the budget.
So I pushed back with a question I didn't know the answer to: why wouldn't we add a new creative, or bump the spend? That question exposed two errors — and neither was a knowledge failure. It knew all of it.
It reused a conclusion without re-deriving it. It had a stored lesson from a different product: don't launch a from-scratch creative with a week left. But this wasn't from-scratch — it was a last-call urgency overlay on the winning creative. Near-zero setup, and it doubles as the close.
It treated high frequency as a verdict instead of checking the disconfirming number. It called 6.77 proof of saturation. But the split around our mid-month landing-page fix was 3.82x after vs 1.47x before. Not saturated — the early days were dragging the average down.
A cached conclusion doesn't sound like an error. It sounds like experience.
So we built "Last call — ends on the 30th," duplicated off the winner, dropped into the same ad set, and held the budget at $50/day on purpose — a fresh creative resets learning and frequency, and buys new people. The last-call ad ran at 9.15x at a $20 CPA. The month closed at 3.12x blended, up from the 2.2x plateau.
Try this prompt (paste before you accept any "leave it alone" call):
Before I accept that recommendation, do three things:
1) List the levers you're rejecting (new creative, new budget, new
audience) and why each is wrong for THIS specific case.
2) Re-derive any lesson you're leaning on — is the current situation
actually the same as the one that lesson came from?
3) Show me the single metric that would prove your conclusion wrong,
and what it currently says.The Job You Can't Automate Away
Here's the trap nobody talks about. We tell everyone to give their AI a memory — and it's right. But memory has its own failure mode: cached conclusions become dogma. The exact system I built to stop the agent from re-deriving everything is the one that handed it the wrong campaign's conclusion and let it apply it with full confidence.
This isn't an ads thing. I run an email agent across five inboxes, and a triage rule that's right for one company gets quietly applied to another where it doesn't belong. Same disease, different room. If you run any AI with a memory, you have this right now — you just haven't caught it, because a cached conclusion is invisible until someone asks why.
What broke it open wasn't my expertise. It was that I was the only one in the room who hadn't already accepted the cached answer.
Everything in its memory is already a conclusion. Your job is to be the one who can still ask the obvious question.
The agent wrote itself a four-step protocol to protect it from its own memory. Steal it.
Try this prompt (paste into your project instructions):
DECISION PROTOCOL — run this before every recommendation:
1) Surface the levers I'm rejecting and justify why, for THIS case.
2) Justify the heuristic I'm using for this specific situation.
3) Check the disconfirming metric — the number that would prove me wrong.
4) Never reuse a conclusion without re-deriving it from current data.Get the system this runs on — free
The whole loop runs on the same rulebook I use for my inbox agent — guardrails already baked in, ready to drop into your own tool. That agent handles about 90% of my inbox silently, not because I built it well, but because it's been corrected hundreds of times by someone who kept asking why. The rulebook is the easy part, and I'll hand you the easy part free.
Rather have it built for you? I build ad and inbox agents for clients. Book a 1:1.
Baldomero Garza — Find me on X, LinkedIn, Instagram, or book a 1:1.
P.S. — If you only do one thing this week: open your agent's memory and ask it why about one stored lesson. Make it re-derive. That's the whole job. Watch on YouTube →