The More I Told My Agent, the Worse It Got — Until I Kept the State, Not the History

The More I Told My Agent, the Worse It Got — Until I Kept the State, Not the History

I spent a week convinced my AI agent was getting dumber. It wasn’t. I was feeding it more, and the extra context was dragging it down. Every conversation I added, every useful snippet I kept in the window, made it slower and less accurate. The fix wasn’t a better model. It was finally admitting that a giant pile of history is not the same thing as knowing where you are.

I wrote about this earlier — how my agent got worse the more I told it. Back then I called it context rot and treated it like an annoyance. This week a Google paper put a number on the exact thing I was feeling: by tracking state instead of replaying the full conversation history, an agent cut its token use by 94% in long sessions — and got more accurate, not less. 65k tokens instead of 1.1 million, with better results. That is not a small optimisation. That is the difference between an agent that thinks and one that drowns.

History is comfort. State is navigation.

When I look at my own vault, I see the same trap. I kept building memory systems that stored everything — every note, every conversation, every half-formed idea — and I told myself that was the point. It isn’t. What my agents actually need is a tight, current snapshot of where we are and what matters next. The rest is noise that makes the signal harder to find.

So I stopped treating my notes as an archive and started treating them as a living map. Delete the obsolete. Keep the state, not the whole journey. It is uncomfortable at first — throwing away a record feels wasteful. But a map that shows you the way is worth more than a shoebox of receipts from every road you never took.

Smaller input. Sharper thinking.

The pattern keeps repeating across everything I build: smaller, high-signal context beats bigger context every single time. The model was never the bottleneck. The context was. If your agent is slowing down or drifting, don’t reach for a bigger model — reach for a sharper state. Cut the history, keep your bearings, and watch the reasoning come back.

What about you — do you keep your whole conversation history, or a lean state of where you are? I’m genuinely curious how many of you are on the archive side.