Glossary · Updated July 2026

What is AI agent memory?

AI agent memory is the durable knowledge an agent carries across sessions — what it learned, what was decided, and what to do differently next time — so it does not start every task from zero. Good agent memory is not a transcript dump: it is retrieval judgment — knowing what to surface, when, and at what depth.

The failure mode of naïve memory is injecting everything into the context window and hoping the model sorts it out. Useful memory does the opposite: it retrieves the few things that matter for the task at hand, at the resolution the moment needs — a one-line reminder here, the full decision record there.

It also separates learned preferences from hard gates, standards, and rules. Preferences are not laws; gates are not suggestions — collapsing the two is how an agent either ignores a real constraint or treats a soft habit as inviolable. Done well, memory is shared: what one agent learns on Tuesday, the right agent knows on Wednesday, without a human re-explaining it.

How it relates to agent management

Memory with retrieval judgment is one of the disciplines of AI agent management. In Vivari it lives in the Vaults layer — living shared memory that agents are born knowing and write back to.