The Four Levels of Agentic Understanding

Your meeting was moved from Friday to Thursday. You saw the update. Your assistant saw it too. Why does its briefing still say Friday?

Because remembering a fact and keeping it up to date are different jobs. I think about them in four steps.

1. Remember

The assistant can bring back something you told it or a document it saw. That beats starting every conversation from zero.

But it can only remember what made it into its store. If you agreed to something in a meeting and never typed it into chat, chat memory alone won't know. And if it has both the old Friday invite and the new Thursday invite, finding both doesn't tell it which one to use.

2. Notice

The assistant has to see changes in the work itself: a moved calendar event, a reply from a customer, a decision in a meeting. Otherwise you're stuck reporting your own life back to it.

Connecting an inbox isn't enough. The system has to notice what happened in it. Even then, it must be careful: an email proves somebody wrote a claim, not that the claim is true.

3. Connect the dots

Imagine a customer says, “Could you send the report by Friday?” You reply, “Yes, I'll send it.” Later, a colleague sends a document. Are those three messages about the same report? Did the colleague's document satisfy your promise? Who was actually responsible?

The words alone don't answer that. The assistant needs to connect the people, the request, your reply and the later action. If it puts the promise on the wrong person, it may have found the right sentence but understood the work badly.

4. Change its mind

When the review moves to Thursday, the assistant should use Thursday. It should still know that Friday was the old date and be able to show what changed it. If the sources disagree and neither settles the matter, it should say so.

The same goes for work that's finished. A completed promise can stay in the history without showing up every morning as something you still owe.

This last step is where many memory systems stumble. They add the new information but leave the old answer standing beside it. Then a search result decides which one you hear.

These four steps aren't a ranking of AI companies. They're questions to ask of any assistant: What can it recall? What changes can it notice? What can it connect? And when the world changes, does its answer change too?

Team0's World Model is built to keep those changes connected to their sources. The point isn't to save more of your words. It's to spare you from repeatedly explaining which of them still apply.