The four-slot context architecture.

The setup

Why your AI keeps forgetting your business.

An operator opens ChatGPT, or Claude, or some other AI tool they pay for. They ask a question about their business — drafting a proposal, writing a follow-up to a client who went quiet, deciding whether to take on a particular kind of job. The answer comes back fast. It is fluent. It is confident. It is also somehow not quite right — too generic, missing the tone they would actually use, citing a consideration that does not apply to this particular client.

They try again. They add more detail this time. Two paragraphs of context. The answer is still off. They close the tab. The whole interaction took six minutes and produced nothing they could actually send.

Almost every time this happens, the problem is not the model. The model is doing exactly what it was trained to do: produce plausible-sounding text given whatever context it has. The problem is that the operator gave it the wrong context, in the wrong place, in the wrong form. "AI doesn't work" is usually "context doesn't work."

"Almost nobody who pays for an AI subscription has been taught to think about context as a separate thing they need to design."

The framework

Four slots. Four homes for the facts.

Every useful fact that helps an AI do your work falls into one of four categories. Each category has a different correct home. Get a fact into the right slot and the AI behaves like a competent specialist. Get it into the wrong slot and it behaves like a stranger who skimmed a memo on the way in.

Here they are in plain English:

That is the whole map. Four slots, four homes. Once you can name them and route a new fact to the right one, you have about ninety percent of what you need to architect context for any AI tool, in any business, at any scale.

Why this matters

The two-paragraph version of what changes.

Three things get noticeably better when an operator starts routing facts to the right slots, before any technology changes.

• The AI stops contradicting itself. Operators often notice the AI gives one answer in one conversation and a different answer in another, on the same topic. That is almost always because the conversations had different ad-hoc context dropped in. When the rules are in a single system prompt and the facts are in a single memory, the answers stabilize.
• Costs stop spiraling. A common complaint is "my AI bills are higher than I expected." Almost always because long context is stuffed into the system prompt and paid for on every single turn. When that content moves to memory or retrieval, you pay for it only when it is actually used, and the bill drops.
• Answers get sharper, not more generic. Counterintuitive: pasting 5,000 words of background at the top of every chat makes the AI worse, not better. The model's attention is finite, and it's all sharing one fixed-size context window — long context dilutes attention away from the question you actually asked. Right-slotted context is shorter and sharper.

That third one is worth saying twice. More context is not better context. Better context is the right facts in the right slots in the right form. The operators who figure this out spend less, get more useful output, and find AI starts pulling its weight in places it had been falling flat.

The trap

The anti-pattern almost everyone hits first.

The most common mistake — the one almost every operator makes the first time someone tells them "AI works better when you give it more context" — is the 5,000-word About My Business document pasted into custom instructions. Industry, clients, team, tools, recent projects, goals for the year, style preferences. Pasted into the system prompt. Felt thorough.

What happens: every conversation gets shallower, not deeper. Bills go up. The AI takes longer to respond. Specific questions get answers that try to address everything in the 5,000 words at once instead of the actual current question. The operator concludes AI is overhyped and starts using it less.

The fix isn't "give it less context." The fix is route the facts to the right slots: identity and rules stay in the system prompt (short), client lists and pricing tables go to memory (retrieved on demand), past proposals and email archives go to retrieval (queried when relevant), and the active conversation handles tool outputs and history (slot four). Same facts, four homes, dramatically better behavior.

There are three more anti-patterns past this one. They are all in the book.

The full playbook

Where to take this next.

Related reading on the site

If you liked this, the next pieces.