LANDMARK · 7 MIN · UNDERSTAND
Memory and state
After this landmark, you can explain where an assistant’s memory physically lives, and therefore what can be inspected, corrected, and deleted.
A model call is stateless: send the same input twice and nothing carries over between them. Everything that feels like memory is built around the model. The conversation continues because the whole history is resent each turn. Long-term memory features work by writing facts to a store — “prefers concise answers”, “works in logistics” — and quietly prepending the relevant ones to later requests. Agent systems keep scratchpads, task state, and notes the same way. This is genuinely good news for governance, because it means the memory is a database rather than something baked into weights: it can be listed, exported, corrected, and deleted. It also sets the failure modes. Memory is only as good as what was written; a wrong fact captured once persists and quietly biases everything afterwards; and because stored items are injected into the window, they compete with your actual request for space and attention.
This is the canonical concept. It stays the same across learner lenses so personalization never changes the underlying facts.
◌ EVERYDAY LENS
What this looks like for you
Open your assistant’s memory settings and read what it has stored about you. Delete what’s wrong or stale — a fact it picked up once will keep shaping answers until you remove it. And treat a shared or work device’s assistant memory as something other people may end up seeing.
MAKE A DECISION
A user says an assistant keeps referring to a job they left two years ago. Where is that coming from, and what fixes it?
CARRY THISOpen the memory or personalisation settings of an assistant you use. Write down one stored item that is wrong, outdated, or that you would not want restated in front of a colleague.