← Blog

Building a useful agent · Memory

How a conversation becomes a memory

July 21, 2026 · 6 min read

Some experiences stay with you for life. Others disappear almost as soon as they happen. We built our agents' memory to work the same way.

Of everything you experience in a day, only some of it becomes a memory. Two parts of your brain decide that. The hippocampus creates the memory, the what, the when, the who. The amygdala decides how strong it is.

For an agent, that experience is the conversations you have with it and the work it does.

Hippocampus
Encode
conversation → separate episodes
Amygdala
Weigh
salience at capture
Neocortex
Consolidate
episodes → facts, nightly
Recall
Resurface
the right memory, per turn
↑ this post
The two lit stages are this post. The dim ones run later, on what capture leaves them. For an overview of the whole system, see the first post.

Conversation becomes memory

The memories you keep of things that happened are episodic memories: what happened, when, and who said it. Someone tells you there's a meeting Wednesday, and that's what you keep. An agent makes the same kind of memory out of a conversation.

It runs in the background, not mid-reply. Once a conversation goes quiet, the hippocampus reads a window of recent messages in one pass. It's a single model call. Out come episodes: one per distinct thing that happened, tagged with what kind of moment it was, each linked back to the conversation it came from.

the conversation
userheads up, the planning meeting is on Wednesday at 2
agentGot it.
userand I'll bring the numbers
episodes
The planning meeting is on Wednesday at 2pm.
The user will bring the numbers to the planning meeting.
↳ each episode links back to this conversation
A short exchange in, two episodes out. One per distinct thing that happened, each linked back to where it came from.

Some memories outrank others

Not every episode counts the same. The agent has your meeting down for Wednesday. Tell it no, it moved to Thursday, and you're correcting a fact it already had, not adding a new one. Capture marks that as its own kind of thing.

That kind is called an act, and who did what decides how much it carries. A correction from you sets the record straight, the highest authority there is. The agent taking back something it said matters too, but that's the agent fixing itself, not you. Yours wins.

Who said it changes what it's worth.

Later on, that's how conflicts get settled. So the tag had better be right.

correction
the user establishes a truth, top authority
decision
a choice made
commitment
a promise or obligation formed
retraction
the agent walks back its own claim
observation
information surfaced
routine
an ordinary lookup or ack
Each act sets a floor the memory can never drop below, whatever else happens to it. Note where the agent's own retraction sits against your correction.

What the memory weighs

How much an episode is worth comes from two things. The act sets a floor: a correction stays heavy even said flatly, a routine lookup starts near nothing. The other is intensity, how forcefully you said it, judged against your own usual tone. A normally terse person who suddenly types in capitals is a spike; the same words from someone who always writes that way are just their normal. The weight is whichever of the two is higher. That's the amygdala's job, and the number it lands on is called salience.

what did they write?

no, it's Thursday, not Wednesday

what kind of act is it?
whose message is it?

ordinary punctuation, occasional emphasis

Floor from the act
Intensity vs this person's normal
the higher of the two wins
final salience
peak
floor wins
Change any of the three and watch the salience move. (Intensity is used only when it's higher than the floor.)

Weight becomes lifespan

A routine lookup fades quickly. A peak moment holds far longer. Salience seeds that lifespan too.

That curve isn't ours. It's a model of human forgetting whose shape traces to 220 million real recall events, logged from people studying vocabulary in a language app. It turns on a single number: how long a memory lasts. We set that number from salience the moment the memory forms.

0%25%50%75%100%time since the memory formed →a peak momenta routine lookupone lifespan
memory lifespan
short-lived
how much is left
67%
retrievable
Higher salience makes a memory last longer. It fades on the same curve human forgetting follows: most of it holds through the memory's lifespan, then a long tail that never quite hits zero. Drag the sliders to reshape it and move through time. Reading a memory never changes it, and time only lowers how easily it comes back, never what's stored.

Every memory fades, but none is ever deleted. Fading only decides what comes back first. When you think about something, or experience it again, a new memory forms, with its own fresh fade. Same with the agent: when something comes up again in a later conversation, that's a fresh experience, and another new memory. The pile of them is what makes a thing stick.

Tell an agent something once, and weeks later it can seem gone. It wasn't erased. It went in weighed low, so it faded, and a passing question might not be enough to pull it back up past everything newer. But give it enough of the context and it surfaces again. It's the same for us, sometimes you just need a bigger reminder before it comes back.

What is actually worth capturing?

Step back to where capture starts. The first question is what's even worth writing down. A language tutor kept correcting one user's Spanish, and capture wrote down every correction. What the agent came away holding was the grammar: moto is feminine, gracias is spelled gracias. For someone learning the language, that's the right thing to keep. But it came from what the model already knew, nothing about the person.

The obvious fix is to record only what the user says and drop the rest. We tried it. It cut a big chunk of the episodes, and a lot of those were things the agent had found out for the user while working. The part that looked like the agent talking to itself was where the memory was. Real, and gone.

What's worth keeping is narrower, and it's about the person: that the user used the wrong gender for moto.

one correction
the user says el mototutor: it's la moto
two ways to remember it
the rule
moto is feminine
the model already knew this
the person
used the wrong gender for moto
worth keeping
The same correction, two memories. One is grammar anyone could look up; the other is about the person.

What capture hands off

Episodic memories just keep accumulating over time, but things we experience multiple times start to become semantic memories, facts. How those facts are created and updated is a job for the neocortex. The only memories that matter are the ones you can recall.

Hippocampus
Encode
done
Amygdala
Weigh
done
Neocortex
Consolidate
episodes become facts
Recall
Resurface
the right memory, in time
↑ still ahead
The map from the top of the post, flipped: capture is done and hands off to the two lit stages. Facts next, then recall.