Build a Personal CRM Claude and ChatGPT Can Read

Your assistant knows your work. Ask it about a project and it holds the thread.

Ask it about a person and it gets vague. It might know your co-founder's name. It probably does not know that Maria moved to Lisbon in March, that she is the one who recommended the book you never read, or that you promised to introduce her to someone six weeks ago.

To be fair, both ChatGPT and Claude now carry memory of their own, and they are decent at it. Mention someone often enough and the model picks up a few facts. That is real and it helps.

It also has three limits worth naming. The memory is inferred, so it records what the model thought you meant. It is hard to audit, so when it gets something wrong you cannot open the list and fix the one wrong line. And it is per vendor, so everything ChatGPT learned about your life stays in ChatGPT the day you open Claude.

A personal CRM fixes all three at once. Not because it is clever, but because you wrote it down.

What a personal CRM actually is

Not a sales tool. No pipeline stages, no deal values, no reporting dashboard nobody opens.

It is one note per person you stay in touch with on purpose. Former colleagues. The person you met at a conference who turned out to be genuinely interesting. Your sister's partner whose job you can never remember.

Each note answers four questions:

• Who are they? One or two lines. What they do, where they are.
• How do you know them? The origin story, in a sentence.
• What did you last talk about? The most recent real exchange.
• What did you say you would do? The promise you made and have not kept yet.

That is the whole format. Four things. If a note takes longer than two minutes, you will stop writing them, and a thin note you actually keep beats a thorough one you abandon in March.

What it looks like in practice

# Maria Okonkwo

Product designer, now at a climate hardware startup. Moved to
Lisbon in March 2026.

## How we met
Conference in Berlin, 2024. We ended up in the same terrible
breakout session and left it together.

## Last conversation
June 2026, voice note. She is finding the new role harder than
expected: strong engineering team, no design process at all.
Recommended "Shape Up" to her.

## Open follow-up
I said I would introduce her to Tomas about design ops. Still
have not done it.

Now open Claude or ChatGPT and ask: "What is the context on Maria before I message her?"

You get the answer. Including the introduction you owe her.

Why writing it down beats letting the model infer it

This is the part that matters, and it is not really about convenience.

You wrote it, so it is right. Inferred memory is a guess about what you meant. Sometimes the guess is wrong, and a wrong fact about a person is worse than no fact. Asking your assistant to draft a message on a bad assumption is how you send something slightly off to someone who notices.

You can read it back. Open the note. That is the whole audit. Every note in Hjarni also keeps its history, so you can see what changed, when, and whether it was you or the AI that changed it. If your assistant writes something into a note that is wrong, you can see the diff and revert it. Opaque memory gives you nothing to revert.

It moves with you. The notes are Markdown in a knowledge base both assistants read over MCP. Connect Claude, connect ChatGPT, and they read the same notes. Switch your daily driver next year and your context comes along. Nothing about this is a bet on one vendor.

Setting it up

The fastest path is the Personal CRM template. It creates the folders and, more usefully, ships the instructions your assistant follows: search the People folder before drafting anything, never invent shared history, and offer to save new details back to the right note.

That last instruction is what makes it self-sustaining. You mention on a Tuesday that Maria got promoted, and your assistant offers to update her note. The CRM maintains itself out of conversations you were having anyway.

If you would rather build it by hand: one folder called People, one note per person, four headings. Start with five people. Not fifty. A CRM with five real notes is useful today, and one with fifty stubs is a chore you already resent.

What to leave out

Do not write anything you would not say to the person's face.

This is a good rule for its own sake, and it also happens to make the notes better. "Struggling with the new role, no design process" is useful context. A private judgment about someone's competence is not context, it is a liability, and it is the kind of thing that reads badly when it surfaces in a draft six months from now.

Keep it to facts and things they told you. That is what makes your assistant useful anyway.

The habit

Message someone. Learn something. Add a line.

That is it. Ten seconds after a conversation, while you still remember. The first month feels like nothing is happening. The sixth month, your assistant drafts a message that opens with the thing they actually care about, and it stops feeling like a filing system and starts feeling like it knows the person.

Because you told it. And you can check what you told it. The personal use case has the rest of the setup if you want it.