Files as Memory

In an earlier post I wrote about config files, the files that tell an AI how a project works. This is about the other side of the same way of working. Not the files that steer the AI, but the files that preserve what we learn and decide.

Because that is what it is, a way of working. Something I apply the moment I develop with AI. Everything I will need again later, I record in a file. Nothing stays in my head, and nothing stays in a chat conversation that will soon be gone. At playitsmart I did it exactly this way. And it works.

Why

An AI has no memory between sessions. But honestly, a human barely has one either. Ask yourself in six months why you made a certain choice, and you will not know. You will remember that you had a good reason. The reason itself is gone.

Knowledge that lives only in someone's head gets lost. Knowledge that lives in files stays.

Which files

Three kinds, each with its own job.

A file for decisions. Every substantial choice goes in, with the reasoning behind it. Not just what we decided, but why. When I had to determine which stocks the system is allowed to trade, it turned out to be three different questions at once. The answer was a list, but the real gain was the explanation of why the list is the way it is. That explanation is fixed.

A file for lessons. Every bug, every mistake, with the fix alongside it. If the same problem ever comes back, nobody has to work it out again.

A file for content. While working I record moments worth a blog post. Not I will write that down later, because later it is gone. Right away, in raw form, in the file. This post partly comes from such a raw note.

Defense in depth

There is a second layer. Every session starts with a fixed ritual: read the current state, the open tasks, the blockers. Every productive session ends with another ritual: update the files with what changed.

I have recorded that ritual in two places, so it does not depend on a single point that can fail. If I forget, the tool catches it. If the tool misses it, I catch it. Two independent safety nets for the same goal.

The takeaway

Treat your memory as the unreliable component it is. Not as storage, but as working memory that will soon be wiped.

This is not an approach I came up with for playitsmart. It is how I work the moment an AI is involved. Everything you will need again later belongs in a file. Decisions, lessons, stories, agreements. A file never complains that it forgot something. Your head does, it just does not tell you.