There is a kind of agent that starts sharp and goes dull. Week one, it answers cleanly. Month three, the same questions come back vaguer, slower, hedged. Nothing in the code changed. What changed is that it had been remembering, faithfully, every day in between.

This is the failure nobody warns you about, because it doesn't look like a bug. It looks like success. The memory system did exactly what it was told — it kept everything. Every passing comment, every one-off question, every detail the user mentioned and never returned to. All of it stored, all of it retrievable, all of it now competing for the agent's attention.

The intuition that gets people here is reasonable: more memory should mean more context, and more context should mean better answers. It's the same instinct that says a bigger library is a better library. But a library is only useful because most of it is not on the table in front of you. The value isn't the volume. It's the selection.

An agent has a fixed amount of room to think in. Every memory pulled into that room to answer a question takes space that something else could have used. When the store is small and curated, the few memories that surface are the relevant ones. When the store is large and undifferentiated, relevance gets diluted. The agent retrieves twelve things, two of them matter, and the other ten are noise it now has to reason around. It doesn't fail loudly. It just gets a little worse, quietly, in proportion to how much it has remembered.

So the real job of a memory system is not retention. Retention is easy — a database does that. The job is triage: deciding, continuously, what is worth keeping at full detail, what should be compressed into a summary, what should be set aside, and what was never worth storing in the first place. A memory system that cannot forget is not more powerful than one that can. It is less.

Forgetting, done well, is not data loss. It is the system deciding that a detail mentioned once eight months ago and never again is not earning its place in the agent's attention — and acting on that decision. Human memory does this constantly and without drama. You do not recall what you had for lunch three Tuesdays ago, and that is not a flaw in your memory. It is the feature that lets you think at all.

The systems that age badly are the ones that treat every memory as equally permanent and equally important. The systems that stay sharp are the ones that treat memory the way attention actually works — a small, well-kept set of what currently matters, with everything else compressed, demoted, or let go.

An agent is not better because it remembers more. It is better because it remembers well. Those are not the same thing, and the gap between them widens every single day the agent is in use.