What an AI agent should remember, and what it should look up

My AI agents keep notes between sessions. They hold my preferences, how a recurring job ran last week, and rules I’ve had to correct more than once. The rule I’ve ended up with is that a live system wins over any note. If a note says one thing and the source says another, the source is right and the note gets fixed.

So I read Anthropic’s write-up on the components of agent memory, published November 6, with real interest. One comparison in it is the clearest illustration I’ve seen of why that rule matters.

Two memory folders

Anthropic had its models play Pokémon Red with a folder of memory files they could read and write. About 10,000 steps in, it compared one run from each of two models. Claude Sonnet 3.7 had written 413 memory files across 225 folders, with a separate file for battles, items and characters it met. It had earned one badge. Claude Fable 5 had written nine files in two folders plus a single top-level dashboard, and had earned four.

The explanation is the useful part. Much of what the older model saved, like the player’s money and position on the map, could be read off the screen every turn. Storing it was redundant, and because the game kept changing, those notes kept going out of date. At one point the model spent 17 steps creating and deleting memory files without taking a single action in the game.

The newer model saved what the screen didn’t show: its current goal, a plan, and rules it had tested. After seven failed attempts to catch one of the easiest Pokémon in the game, it wrote down the timing change that finally worked.

These are single runs in a game, so I wouldn’t lean on the numbers. Commenters on Hacker News pushed back on calling the review process “dreaming,” and that’s fair. The distinction between remembering and looking up holds regardless of what you call it.

What this looks like in crew travel

Take a hypothetical agent that books hotel rooms for airline crews during a disruption.

Some facts change by the minute and have an owner: room availability, contracted rates, which crew members are legal to fly tomorrow, the current schedule. They live in systems of record. The agent should query them every time it needs them and never trust a copy from an earlier session. A remembered rate or room count is the stale Pokémon note: plausible, and wrong.

Other facts exist only because someone learned them the hard way. Say one property near an outstation only confirms late-night arrivals by phone, so an email confirmation after 10 p.m. isn’t really a confirmation. No booking system records that. A good coordinator carries it in their head, and it’s exactly what an agent’s memory is for.

That kind of note needs two more things: how the agent learned it, and when. “Learned from three unconfirmed late arrivals in March” is something a person can check. “Hotel needs a phone call,” with no source, is a rumor that will outlive the hotel’s next change of management.

Memory needs review

The write-up has a second example worth knowing. A model wrote down the wrong location for a building, trusted that note in later sessions, and kept walking into the same trap. It recovered only when a review pass checked the note against the original session transcript and corrected it.

In operations, the transcript is the booking record, the confirmation and the call log. I’d give a person a regular review of what the agent has written down, with each note’s source attached and the authority to delete it. It’s the same idea behind how I think about tasks for AI agents: a record should carry its own evidence. And when a remembered rule turns out to be fully deterministic, it may not belong to the model at all. It belongs in code, as I found when a short policy file beat the model I was paying to read it.

The test

Before you give an agent memory, list what you expect it to save. For each item, ask one question: is there a system that already knows this? If there is, the agent should look it up. If there isn’t, it should write it down, say where it came from, and expect someone to check.