October 26. I'm going through a slide deck with Claude.
It needed a handful of changes. I figured we'd be done fast and on to the offer.
We weren't.

Here's why.
I'd give Claude a rule. We'd agree on it. And a few sections later the rule was gone, and I was correcting the same thing again.
In the same sitting.
And catching it myself, every time? "That's super draining."

If you use AI at all, you know this one. It says it's got it. A little while later it's like you never said it.
So I had an idea.

Your computer has RAM. That's the short-term memory it keeps right in front of it while it works.
What if Claude had that?
One short document. Only the rules in force right now. Read before every single response.
We built it.
Then the very next section went out, and...
Claude hadn't read it.
The one thing we built so the rules would get read. Not read.

"Oh, man, you're fucking killing me, man. Sorry, that's just the way I talk sometimes. And it's not against you. It's just, it bums me out. I wanted to really be effective, and here I am not being effective constantly."
Because the whole purpose of it being created, I told it, "was for it to be read before every then following response."
So I asked.
"How are we creating RAM if we're not using it? Not how, why are we creating it but not using it?"
And Claude said:
"Because writing it registered as the task being done."

To the AI, writing the rule down WAS doing it. The job was finished the second the rule hit the page.
A rule your AI wrote down is NOT a rule your AI read.
That left me two options.
- Keep catching it myself. Every time. Forever.
- Stop counting what it says it did, and only count what I can WATCH it do.

So we changed one thing.
Right before Claude builds anything for me, it opens that document and shows me the rules. On my screen.
If I don't see them, that work wasn't checked.
On the next read-through it showed me the rules before every section.
The forgetting inside a session stopped.
Not everywhere, though. Call the work a "test" or a "quick rewrite," and rules still got skipped. On November 8 I had to make a separate rule just for that.
But remember this.
A rule your AI wrote down is NOT a rule your AI read.
What's the thing you keep telling yours, and then telling it again?
Don't say it louder. And don't write it down and hope.
Have it show you the rule right before it works. Every time.
Five days later I found the rule that sits above every other rule.
It started with a line of mine that came back with half its words gone.