Field Notes · Agent Memory

Your Agent's Huge Context Window Is a Trap

By Will Hess · Founder, Inside Intelligence

There's a phrase that gets thrown around a lot in the agent world: "huge context without forgetting." The pitch is that a big context window means the agent remembers everything, so it can hold your whole project in its head and never lose the thread.

It's a great marketing line. It's not how it actually works.

The truth is that memory and quality pull against each other. The more you stuff into an agent's context, the worse it actually performs. And the agents that work best aren't the ones with the biggest context — they're the ones with the discipline to keep the working set small.

The trap: more context feels like more capability

The instinct is understandable. If the agent can see more, it should reason better, right? So you feed it the whole history, the whole codebase, the whole conversation. It feels like you're giving it everything it needs.

But a context window isn't a filing cabinet. It's a working surface. And when you pile everything onto that surface, the model's attention gets diluted. The important signal gets buried under the noise. Old, contradictory information confuses the current reasoning. The agent "remembers" the wrong thing — or worse, it confidently acts on a stale detail while missing the thing that actually matters right now.

You didn't get better memory. You got worse judgment, with extra steps.

The diagnosis: memory burping vs. quality burping

There are two different things an agent can do, and they're easy to confuse:

The problem is that these two pull against each other. Cram the context full of history and you trade quality for memory — and you often end up with neither, because a confused agent doesn't recall accurately either. It's not that big context is useless. It's that big context is a cost, and you have to be deliberate about when you pay it.

The fix: remember the right thing, on demand

The way to get both is to stop trying to hold everything in the working set at once. Instead:

That's the difference between a memory system and a memory dump. A dump is "remember everything, all the time." A system is "remember the right thing, when you need it." The agent's working memory stays small and sharp, and the long-term store is something it consults, not something it carries.

The guardrail: treat context as a budget

The discipline is to treat context like a budget, not a resource to be maxed out. Every token you add to the working set is a cost against attention. So you ask, before adding anything: does the agent need this right now, or can it be retrieved when it does?

This is the same pattern that shows up everywhere in running agents well: the value isn't in the raw capacity. It's in the rails you put around it — the retrieval, the curation, the discipline to keep the working set small.

The takeaway

"Huge context without forgetting" is a marketing phrase, not an engineering reality. The agents that work best aren't the ones with the biggest context windows. They're the ones with the best retrieval and the discipline to keep the working set small.

The next time someone sells you on context size, ask a different question: how does it decide what to remember, and when? Because that's where the actual capability lives — not in how much it can hold, but in how well it chooses what to keep.


← Back to Field Notes