What to put in your company's AI memory first

Say you are convinced: your AI needs your company's knowledge, not just a clever prompt. Good. Now comes the question that stalls most people. Where do you even start? The knowledge is enormous and scattered, and staring at all of it is how the project quietly dies.

The answer is that you do not start with all of it. You start with the handful of documents that change the most answers, and you leave the long tail for later. Here is that ordered list.

Start with who and how

Before any document about the business, write down two things about the people using the AI.

• Who you are. Your role, what you produce, how you write. One page. The fastest way to create it is to ask the AI to interview you and write it up. This single file is the difference between generic output and output that sounds like you.
• How the company works at the top level. What you do, for whom, your tone of voice, and the rules the AI should always follow. In Claude Code this lives in CLAUDE.md and is read automatically every session.

These two are first because they touch every answer, not just answers about one topic.

Then the five documents that pull their weight

After the who and how, add the business documents in this order. Each one earns its place by how often it changes an answer:

1. What you sell. Products or services, plainly described, with the positioning. The AI cannot talk about your business if it does not know what the business is.
2. Pricing. The single most-asked-about thing, and the one where a wrong answer is most expensive.
3. How you win. Your sales playbook or the way you actually pitch: objections, differentiators, the arguments that land.
4. Who you are up against. A competitor rundown from public sources, so the AI stops comparing you to the wrong companies.
5. Proof. Case studies and results, described without naming clients you have not cleared to name.

If you only ever wrote these seven things down, your AI would already answer better than 90% of the "AI for business" setups out there.

What to leave for later

The instinct to be complete is what kills the project. Deliberately skip, for now:

• Anything sensitive: contracts, real customer data, internal financials. That belongs behind stricter controls, not in a first draft of your knowledge.
• The long tail of edge-case documents. You can always add them when a real question exposes the gap.
• Perfect formatting. Plain, readable text beats a beautiful document that took a week.

A good rule: add the next document only when a real question reveals you needed it. Let actual use, not imagined completeness, drive what you write down next.

A weekend, not a quarter

Here is how the starter set fits into two afternoons, not a project plan.

Saturday morning, do the two personal files: ask the AI to interview you for your own profile, then write the top-level description of how the company works. An hour, maybe two. Saturday afternoon, knock out what you sell and your pricing, the two documents that answer the most questions. Sunday, add the sales playbook, a competitor rundown from public sources, and a couple of proof points. That is the whole starter set, and it is enough to change every answer you get for the next month.

Notice what you did not do: you did not try to be complete, and you did not polish. Both are the traps that turn a weekend into a stalled quarter.

Common questions

How long should each document be?

Shorter than you think. One readable page per topic beats ten pages nobody maintains. The AI does better with a clear page than an exhaustive one, and you are far more likely to keep a short document current.

What if two documents disagree?

Then you found something valuable: a decision nobody made. Pick what is true and fix the other. This is exactly why knowledge should be plain, readable markdown you can correct, not a black box you cannot inspect.

Do I have to write it all by hand?

No. Point the AI at material you already have, old proposals, your website, a pricing sheet, and have it draft each document, then you correct. Converting what exists is faster than writing from scratch, and fixing a draft is easier than facing a blank page.

What happens after the starter set?

You let real questions pull in the next document, and eventually the knowledge needs to grow on its own as work happens. That is the jump from a static base to a living one: from meeting to memory.

The takeaway

The starter set is small on purpose: who you are, how the company works, and five business documents. That is a weekend, not a quarter, and it is enough to flip your AI from generic to grounded.

The harder question is what happens after the weekend, when this knowledge needs to be shared across a team, kept current as things change, and reachable from any AI client instead of one laptop. That is exactly the jump memrelay is built for. The full guide to building the personal version first is in the practical Claude Code guide.