Give your AI a memory that lasts between chats

Every new chat with an AI starts blank. It doesn't remember what your best client said last month, which offer you dropped, or why you price the way you do. For a while I dealt with that by pasting the same background in at the start of every session. It worked, and it wasted a lot of time.

What fixed it was giving the AI a set of notes that it writes and keeps up to date itself. When I asked whether that meant some complicated "vector database", the honest answer was no. It's a folder of plain text files and a habit. This guide shows you how to set one up.

The idea in one paragraph

The pattern I use comes from the AI researcher Andrej Karpathy, who describes it as an LLM wiki. (An LLM, or large language model, is the kind of AI behind Claude and ChatGPT. A wiki is a set of linked pages.) You keep three things. Raw sources: transcripts, emails, meeting notes, never edited. The wiki: short pages the AI writes from those sources, one per topic. The rules: how pages get written and kept tidy. The AI does the bookkeeping, which is the part people give up on when they keep notes by hand.

Step 1: Know which memory goes where

Not everything belongs in notes. I keep four kinds of memory, and each has one home:

• Stable rules go in your instruction file (CLAUDE.md, covered in the previous guide). Who you are, how to work, what never to do.
• Knowledge goes in the wiki: what you've learned and decided, in sentences.
• Records go in a spreadsheet or simple database: anything you'd filter, sort or add up, like leads, tasks or hours.
• What happened today goes in the chat itself and can be forgotten.

The common mistake is putting records in notes. A list of 40 leads written into a document can't be sorted, and the AI will miscount it. Keep notes for prose and tables for numbers.

Step 2: Make the wiki folder

Inside your top business folder, create a folder called wiki, or one inside each client folder if you work with clients. Then make two files:

• `index.md`, a list of every page with one line saying what's on it.
• `open-questions.md`, everything you don't know yet.

Then add pages by topic, not by date. "Pricing", "Ideal customer", "What clients complain about", "Services we stopped offering". Each page is short and written in plain sentences.

Step 3: Feed it as you go

The rule that makes this work: every time a new source comes in, update the wiki in the same sitting. A client call, an email with answers, a decision you made on a walk. Save the raw source untouched, then ask the AI:

"Read this transcript. Update the relevant wiki pages with anything new. Quote the client's exact words where it matters. Move anything that's now answered out of open-questions.md, and add anything new we still don't know."

If you skip this step, the wiki slowly turns into a diary. That happened to me. I had hundreds of dated session notes and almost no pages by topic, so nothing was findable when I needed it. Write pages by topic, and keep the dated notes as raw sources.

Step 4: Read the index first

When you start a session, the AI doesn't need the whole wiki. It needs the index and the open questions, which together are usually short. From there it opens only the page it needs.

Put a line in your CLAUDE.md like: "Before writing anything for a client, read their wiki/index.md." That one instruction is most of the value.

Step 5: Use the open-questions page as your to-do list

This turned out to be the most useful file of the lot. Because every unknown lives in one place, it becomes the list of things to ask next. When I prepare questions for a client call, I build them only from that file.

I learned that the hard way. One agenda was built from an old list of interview questions, and it was full of things the client had already answered. A list of what you once asked is not the same as a list of what's still open.

Step 6: Write the words people will search for

Most simple search tools match the words actually on the page, not the meaning. If a page says "rates" and you search for "pricing", you may not find it. When the AI writes a page, ask it to use the real names: the client's company name, the product name, the exact phrase a customer used.

Step 7: Tidy it now and then

Once a month, ask the AI to check the wiki: pages that contradict each other, facts a newer source replaced, quotes that turn out to be a common saying rather than something your client actually said. Ten minutes keeps it trustworthy.

One more thing the wiki can't do: it only records what was said. If an important topic never came up, there's no page for it, and nothing will warn you. Read your own index occasionally and ask what's missing.

What to do next

Start small. One wiki folder, an index, an open-questions file, and three pages about the things you explain most often. Feed it after your next client call and see how much less you repeat yourself.

Next in this track:

• Plan before the AI builds anything, using what's in the wiki so the plan starts from facts.
• Lock down your accounts before you connect AI, before it starts reaching your email and files.

If you'd like help deciding what belongs in your AI's memory and what doesn't, ask me for the free audit. I'll look at how your business knowledge is stored now and where the gaps are.