From: Tom
Somewhere with good WiFi and better coffee
Saturday, 11:42 a.m.
The same mistake. Again.
My copywriter agent used an em dash. I’d told it not to. Twice.
I fixed it. Published the post. Moved on.
Three days later: another em dash!!!!?!?!?!?!?!?!?!?!!??!?!?!?!F*$#%^$K!!?!??!?!
That’s when it hit me. My agents weren’t learning. They couldn’t. Every session started fresh. No memory of the last conversation. No record of my corrections. No improvement.
Seven smart agents. Zero collective memory.
--
The Problem Nobody Talks About
AI agents don’t remember.
Not because they’re stupid. Because that’s how they work. Each session is a blank slate. The context from yesterday? Gone. The feedback you gave last week? Evaporated.
Here’s what that looks like in practice:
Week 1: Gatekeeper rejects 40% of drafts. Same issues keep appearing. Voice inconsistencies. Wrong formatting. Patterns I’ve corrected before, surfacing again.
I was the memory. The human hard drive holding all the preferences, patterns, and “please don’t do that again” moments.
It wasn’t sustainable.
---
The Fix That Changed Everything
I created a file called `learning-log.md`.
Simple idea: one document that captures what works, what doesn’t, and what we’ve learned. Every agent reads it before working. Every review session adds to it.
The structure has three sections:
Active Patterns (Apply These Now)
What works. The moves that get approved. Specific examples of good decisions.
Example entry:
| Do This | Not This |
|---------|----------|
| Start with a specific moment | Start with a generic statement |
| Use short punchy lines | Write long flowing paragraphs |
Common Mistakes to Avoid
What keeps going wrong. The patterns that trigger rejection.
Example entry:
1. Using em dashes (—) anywhere. Use periods or commas instead.
2. Opening with a lesson instead of a moment
3. Writing headlines that could apply to any competitor
Iteration Log
What we learned from each review session. Timestamped. Specific.
Example entry:
## 2026-01-17 - Copy Feedback
### What Worked Well
- Spaceship metaphor was instantly graspable
- Single-word CTA “loop” converted well
### What Needed Improvement
- Opening was too abstract. Start with action.
### Pattern Discovered
- Value-first posts outperform direct offers on reach
The Flywheel Effect
Here’s what happens now:
1. Agent creates content
2. Gatekeeper reviews and scores
3. Patterns get logged (what worked, what didn’t)
4. All agents read the updated log
5. Next content is better
Week 1: 40% rejection rate. 10 minutes per post.
Week 4: 15% rejection rate. 2 minutes per post.
The system learns. The agents improve. The quality compounds.
Most people use AI the same way forever. Same prompt, same quality, same frustrations.
We built a loop that gets tighter every time.
---
The Deeper Point
Your AI agents are only as good as their instructions.
But here’s the thing: instructions can get better automatically.
Every review is a chance to teach. Every mistake is a pattern to log. Every success is a move to repeat.
You don’t need to remember everything. You need a system that remembers for you.
That’s the learning log.
That’s it for this week.
If this was useful, forward it to someone (real human) building with AI.
See you next week ✌️
(the guy whose AI team remembers everything except birthdays)
P.S. This newsletter was 93.7% made by my AI team. I stayed in the loop.
P.P.S. Recommended read: How 7 AI agents speak with one voice. (Hint: it involves three files they all read.)
P.P.P.S. Want to build your own AI team? I teach the full system in a 90-minute online workshop. DM me or reply “log” to get the details.
P.P.P.P.S. I read every reply. The real me 🤓