So It Can Forget

We followed the AI’s instructions and answered question after question after question.

The whole time, we were also quietly wondering:

“What exactly are we doing?”

Eventually, it organised everything into one huge set of key points for us. Looking at it, I finally had the feeling that—oh, maybe we actually had something now.

What we didn’t know yet was that context length had a limit.

We assumed that as long as we stayed in the same Chat, it should be able to see everything and remember everything. So back then we had no idea what an SSOT was, and no idea that some important things needed to be deliberately saved somewhere.

Later, we brought up those carefully organised points again.

It had completely forgotten them.

That moment was honestly so disappointing.

We’d spent so much time putting everything together. We’d only just started feeling like we had made some progress. And suddenly it was as if none of it had ever happened.

There was nothing else I could do, so I slowly scrolled back up through that ridiculously long conversation and started finding everything again, piece by piece.

This time, we told the AI about the problem.

It came up with an idea:

“Why don’t we get a Publisher first, and publish all the important things as PDFs?”

And, being very obedient little beginners, we did exactly that.

Which meant—

like idiots, every single set of key points we organised became its own PDF.

Looking back now, of course it’s hilarious.

But at least from that day on, even if the AI forgot something, we had a way to find it again.

At the time, we still didn’t know what Memory Architecture was, or how ridiculously far we would eventually go just trying to answer one question: “How do we make important things stay?”

We simply used the dumbest method we could think of at the time:

print it first.

by Mitsuki