Do AI agents actually need long-term memory?

I've been thinking a lot about what happens when AI agents move from short tasks to workflows that last days or even months.

Longer context windows help, but I'm not convinced they solve the whole problem.

For example, imagine an agent working with the same customer for six months.

It might need to remember that:

• a proposal was already rejected

• a customer's priorities changed

• an old decision was replaced by a newer one

• a promise made three weeks ago is still unresolved

Simply putting more conversation history into the context window doesn't necessarily tell the agent what's still true or what matters now.

This makes me think the harder problem isn't storing more context — it's maintaining useful state over time.

I'm curious how other builders are approaching this.

Do you think long-term memory will become a standard layer in AI agent architecture, or will bigger context windows eventually make dedicated memory systems unnecessary?