The Three Intelligences

Dylan Telano · Oct 23, 2026 · 2 min read

TL;DR: Long-running AI needs three things: reasoning, memory and state. We scaled the first two. AI agent state got skipped, and that is why agents fall apart.

AI agent state is the missing third

Every long task an AI does well needs three kinds of intelligence. Reasoning, to work out the next step. Memory, to recall what it has seen. And state, to know what is true right now.

The industry has poured almost everything into the first two. Models reason better every quarter. Context windows went from thousands of tokens to millions, and memory layers bolt on the rest. State, the third, mostly got skipped.

A useful way to picture it

If the model is the CPU and the context window is RAM, state is the file system. It is the durable record of what happened, what still applies and what comes next. Every serious computer has one. Most AI agents do not.

We built a fact-losing factory

Without state, an agent rebuilds the world from a pile of tokens on every turn. Early in a task that works. Two hundred steps in, it does not. A price changed on Tuesday. A requirement moved from v1 to v2. A customer was promised a replacement. The facts are all in there somewhere, next to the old versions they replaced.

So the agent redoes finished work, quotes last month’s policy, or confidently builds on a requirement that no longer exists. Bigger models make this rarer, not impossible. More memory makes it worse, because there is more stale material to get wrong.

Externalize the state

Our bet is simple. State should not live inside the model or the prompt. It should be kept outside, as its own layer, and managed across its whole life: observe what happened, verify it against evidence, reconcile what is true now, and hand the current state to the model before it acts.

That is what Cricket does. The model still reasons. Memory still recalls. Cricket keeps track of what is true.

Reasoning tells an agent how. Memory tells it what was. State tells it what is.