What does “state” actually mean in modern AI tools — and are we using it wrong?

Most AI tools still feel stateless unless we manually paste context or rely on limited memory features.

But in reality, every meaningful workflow needs state:

-past decisions
-user preferences
-intermediate outputs
-evolving context

Without it, we’re constantly restarting from zero.

So the question is — what should “state” look like in AI tools?

Should it be:

-conversation-based (chat history)?
-structured (knowledge graph, database)?
-or something hybrid (semantic memory + threads)?

And more importantly — who controls it: the user, the tool, or the system layer in between?

Feels like this is one of the biggest unsolved UX problems in AI right now.