Why does your AI keep making the same mistake?

Dangerously good memory

I wanted an AI that remembered everything. Then I discovered that the most dangerous AI memory is not an empty one. It is a memory that remembers you confidently and incorrectly.

One small false assumption

My personal AI workspace is called Sofia. At one point, she made a perfectly logical assumption: if I had sent her a new message, I must have read her previous answer.

I had not. I often write several thoughts in a row and read the answers later. That small mistake could have made Sofia offer me work that had already been completed in the meantime, or connect my thought to an answer I had not yet seen.

A minor detail? Once, yes. By the third time, no.

A new prompt is not learning

We usually correct AI with a new prompt: “No, I meant this instead.” The answer improves and we move on. The following week, we start again from the same mistake.

That is not learning. It is like showing a new colleague where the coffee machine is every single morning.

A better prompt gives you a better answer today. A better correction can give you a better partner tomorrow.

45 changes, 23 rules

During one 12-day experiment, the file containing Sofia’s most important working rules received 45 Git changes. By the end of that period, 23 active rules remained.

Those were not 45 revelations from a mountaintop. I refined some rules several times, merged others and later deleted a few. It is closer to sculpting: shave a little off here, remove one false assumption there, then see whether the next answer actually improved.

An AI Partner is never truly finished. But it can step on the same rake less and less often.

A simple correction format

My simplest correction format is: “When X happens, do Y, because Z.”

For example: “When Tauri sends a new message, do not assume he has read the answers in between, because his writing queue and reading queue move independently.”

Then I ask AI to repeat the rule in its own words. I add it to the permanent working instructions, and the next similar situation becomes a test. If the same mistake returns, the rule was not good enough — or I corrected the wrong problem.

If your AI tool does not keep persistent instructions, create a correction document and attach it at the start of important work. It is less sexy than a “magic superprompt”, but works far better than sighing about the same thing for the tenth time. :)

Memory must learn to forget

Corrections cannot simply accumulate forever either. Uncorrected AI memory resembles an office wall where every former manager’s sticky note still has the force of law.

If I walked into the kitchen with muddy shoes once, that does not make me “a person who always walks into the kitchen with muddy shoes”. Good memory must distinguish a one-off incident, a recurring pattern and knowledge that is no longer current. Sometimes AI does not need another lesson. Sometimes it needs help forgetting one.

Context does not grant rights

Inside a company, the same question becomes even more important. Giving AI the entire company Drive is like dumping a library on a new employee’s desk and saying: “Now you know us.”

It does not. It cannot automatically tell which document is current, whose opinion it contains, what was a temporary thought or what it is even permitted to use. More context gives AI more capability, not more rights.

That is why I would start with one real task, one recurring error and one correction whose disappearance can be checked. Only then would I add the next task.

Measure the mistakes that disappear

Do not measure only how many files AI has read or how clever its answers sound. Measure how many recurring mistakes disappeared, how many false assumptions surfaced and how often you had to explain the same thing again.

Next time AI gets something wrong, do not immediately press regenerate. Ask: “What persistent rule should you learn from this correction for next time?”

An AI Partner is not born when it gives a brilliant answer. It is born when your “no, that is not how it works” leaves a trace in the system.