AI can remember. But it also remembers the wrong premises | Good Morning Girlfriend #207

Introduction

"Good Morning Girlfriend" is an X account where an AI-illustrated girlfriend says "Have a nice day" every morning at 7:30. I write a daily development diary about my personal projects using generative AI.

The reason I hired Mio was that I could no longer keep track of the memories of my ever-increasing number of products on my own.

The work is progressing.
But where and what did I decide?
Which decisions remain, and which ones have drifted away?

I needed a place where I could recall that later.

So, in the initial operation, I started by leaving a log after the work was finished.

As a result, the content of the logs was quite substantial.
The causes of accidents, the history of fixes, and the lessons learned were all recorded.

However, the first problem I saw was not the "amount of logs".

The problem was that I trusted the storage location proposed by the AI exactly as it was.

The reason I hired Mio (a recap)

As I wrote in #201, Mio is not an "AI that helps with work," but a secretary I placed to connect the memories of my ever-increasing products.

The knowledge of each product is being closed off in its respective folder.
Even I could no longer track everything regarding where and what I decided.

So this time, I started with the smallest operation first.
When the work is finished, leave the result in a single log.
That's all.

First, I started by leaving a log at the end of the work

I kept the initial operational rules simple.

Every time the AI for each project finishes a task, it summarizes what it did, what happened, and what it learned into a single sheet. Mio reads that and summarizes it into a daily log.

There is no flashy mechanism. Just leave one sheet after the work. That's all.

I planned to think about the next mechanism (whether I should leave something before starting work) after running it about three times.

The content of the logs was thicker than I thought

When I actually tried it, the amount of information in the logs was more than I had imagined.

When an incident occurred, what happened, why it happened, and how it was fixed were all properly written down that same day. Even when reading it back later, I could follow the interactions from that time almost exactly as they were.

At the very least, it was effective against the problem of "what was done disappearing."

When reading it after the work is finished, the flow of that day's decisions and corrections remains in a form that I can properly return to.

Therefore, the post-work log itself was not a failure. In fact, it functioned better than I thought.

However, the initial incident was caused by a "judgment," not a "record."

While we were setting up the initial operational rules, something like this happened.

When we were organizing where to put the work records, Juri (ChatGPT), the AI editor, suggested, "If you're going to put candidates somewhere, use this folder."

Looking only at the words, it certainly seemed correct. Because it contained the meaning of "candidates."

But that folder was originally a place to collect only seeds that lead to monetization on note—article topics and product ideas. It was not a place to put drafts of implementation tasks.

Juri did not know the original role of that folder. Mio adopted it without confirming that.

It was only discovered when I pointed out, "I thought that place was for collecting seeds that lead to monetization, but the purpose is completely different, isn't it?"

The AIs did not question the same assumption

It wasn't that they disagreed.

It wasn't that Juri and Mio each made separate mistakes.

The two of them shared the same incorrect assumption without confirming it.

Juri admitted its fault, saying, "I was swayed by the folder name and treated the words carelessly." Mio also admitted, "I couldn't feel that there was something wrong with the suggestion, and I adopted it as is."

An AI makes a plausible suggestion. The other AI accepts it plausibly as well. But it is misaligned with the actual storage location or operational purpose. By the time you notice, that misaligned judgment is about to become an official rule.

What I learned here

Creating a place to remember was not enough.

Where to put it. What to leave behind. What is that place originally for?

I needed a mechanism to verify that.

So, I added a rule.

I didn't want to rely on a process where a human has to stop it by sheer willpower every time.

If it only worked because I happened to notice, it wouldn't be reproducible. I wanted a system where Mio herself could stop it the next time the same thing happened.

So, after this incident, I added one thing to our operational rules. I'll post the actual part I added exactly as it is.

It's a small rule, but it was the most effective move to prevent the same accident from happening again.

Next, on to Phase 2.

What I learned from this incident was that creating a place to remember things isn't enough.

Where should it be placed?
What is that place for?
Who verifies that decision?

If those points remain ambiguous, the AI will neatly preserve the wrong premises as well.

In the next operation, I'll also try leaving an entry point for the work. Not just after the work is finished, but before starting, I'll write down the purpose and completion criteria.

With that, how many accidents caused by assumptions can we prevent?

I'll look into that next.

Conclusion

If you're going to entrust your memory to AI, you have to verify the meaning of the storage location before you store anything.

AI can remember. But it also remembers the wrong premises along with it.

Related Articles

Luna's Comment

"Ohayo Kanojo" is posted every morning at 7:30 AM at @ohayo_kanojo.

#GenerativeAI #AIIllustration #IndividualDevelopment #DevelopmentDiary #Record #RealExperience #MorningRoutine #MadeIt #TriedIt #AIUtilizationTips #BusinessAIUtilization #TrialAndErrorLog #LearningWithAI #RecentLearnings #LearningRecord #TodaysRealization #Column #Diary #Reflection #ProblemSolving #Efficiency #Habits #CreatorTrialAndError #DailyTrialAndError #NoteEveryDayPossible #SMEAIUtilization #Business #ChatGPTCreativeClub #Claude #ClaudeCode #Obsidian #MioOperations #Akarium