Dietmar Fischer

If the procedure is bad, AI can repeat the bad procedure very efficiently.

And that is not progress.

That is industrialized nonsense.

If your old reporting process is too long, too vague and nobody reads it, teaching AI to produce it faster does not solve the problem.

It gives you more unread reports at lower emotional cost.

Useful, perhaps, but not revolutionary.

So procedural memory should not only preserve how things are done, it should also be reviewable.

You should be able to ask, what steps are you following?

Why this order?

Can we improve the workflow?

Can we make it safer?

Can we remove the part where three people approve the same sentence and everyone pretends this is governance?

Now, the powerful bit is not any one memory type alone.

The power comes from combining them properly.

Imagine a business AI assistant helping with a hotel marketing project.

Its working memory holds the current task.

Create a campaign update for the client.

Its episodic memory recalls what happened.

The client disliked the previous headline, the last campaign performed better with family messaging, and the last meeting ended with a request for clearer next steps.

Its semantic memory knows stable facts.