Dietmar Fischer

Recap.

Memory is not one big digital bucket.

So let us pull the main idea together.

AI memory is not just about making a chatbot remember more stuff.

That sounds useful, but it can quickly become messy, risky, and slightly creepy.

The smarter idea is to separate memory into different types, because not every piece of information has the same job.

Working memory is what the AI needs right now.

It is the current task, the current instruction, the thing on the desk.

If you ask for a short client update, the AI should remember that it is writing a short client update, not suddenly wander off and produce a strategic manifesto with 12 headings and the emotional weight of a tax audit.

Episodic memory is what happened before.

It is the project diary.

The client rejected a headline.

The team chose version B. The last report was too long.

These memories help the AI keep continuity, but they should not automatically become permanent rules.

One awkward meeting does not mean the entire client relationship is doomed.

Sometimes people are just hungry.

Semantic memory is stable knowledge.

These are the facts the AI should treat as generally true.

Company policies, brand rules, product details, customer segments, legal limits, pricing rules.

This is where accuracy matters.