How does AI memory work?

AI models don't remember on their own. Memory is something built around them. Here's how, without jargon.

Updated

Short-term: the conversation

A language model only sees the text it's given for each reply — roughly the recent conversation. When that falls out of view, it's gone, unless something saved it.

Long-term: saved memories

Good systems store important things separately: facts about you, shared moments, the person's own background. Before each reply, the relevant ones are looked up and included.

Choosing what's relevant

Remembering everything isn't the goal. The system needs to pick what matters for this moment so the conversation stays focused.

What makes memory trustworthy

Separating what they know about you from their own story, saving only what you actually said, letting corrections win, and letting you delete it all. Those rules matter more than raw capacity.

Read next