Give Your AI a Memory

Key takeaways

• Memory is plain text files your AI reads before it starts: CLAUDE.md for Claude, AGENTS.md for Codex.
• Four kinds: who you are, how you work, what you are working on, and where things are.
• Let your AI interview you and draft the files, then fix what is wrong.
• Every time you correct it, say "remember that," and keep one fact per file.

The biggest difference between people who get value from AI and people who don't isn't the model. It's context.

Every time you open a fresh chat and explain who you are, what your business does and how you like things written, you're paying a tax. Most people pay it so many times they give up.

Memory ends the tax.

What memory is

Memory is a set of plain text files your AI reads before it starts work. Nothing fancy.

Claude reads a file called CLAUDE.md. Codex reads one called AGENTS.md. The setup guide makes both, and they say the same thing.

Think of it as the onboarding doc you'd give a sharp new hire on day one.

 ~/ai-system/
 ├── CLAUDE.md          ← read every session (Codex: AGENTS.md)
 └── memory/
     ├── MEMORY.md      ← the index: one line per file below
     ├── me.md          ← who you are, your business, your voice
     ├── rules.md       ← how you want things done
     ├── people.md      ← team, key clients, who's who
     └── projects/      ← one file per thing you're working on

The four kinds of memory

Who you are. Your role, your business, your customers, what you sell and for how much. How you write. The words you use and the ones you'd never use.

How you work. The rules. "Always ask before sending anything." "Point form, not paragraphs." "Never book calls before 11am." Every time you correct your AI, that correction belongs here so you never make it twice.

What you're working on. Active projects, goals, deadlines, who's involved. These go stale, so date them.

Where things are. Which Drive folder holds the contracts. Which login is for what (the location, never the password). The dashboard you check on Mondays.

The four kinds of memory

Kind	What goes in	Example
Who you are	Business, customers, prices, voice	"Our minimum job is $5,000."
How you work	Rules and every correction	"Point form, not paragraphs."
What you are working on	Projects, goals, deadlines, dated	"Q4: launch the new offer."
Where things are	Folders, dashboards, where logins live	"Contracts are in Drive > Clients."

How to fill it

Don't sit down and write it. You'll hate it and it'll be generic.

Let your AI interview you instead. The AI OS Starter Kit has the interview prompts and templates. It'll ask one question at a time, pull what it can from your email and Drive, and draft the files. You fix what's wrong.

Then the most important habit in this whole book:

When you correct your AI, tell it to remember.

"That's too formal. Remember that." "Don't cc Sarah on these. Remember that." "Our minimum is $5,000. Remember that."

Every correction you save is a mistake it won't make again. After a month, it writes like you, knows your people and follows your rules. That's when it starts to feel like a team member instead of a tool.

Keep it clean

Memory gets messy if you let it.

• One fact per file. Easier to update, easier to find.
• Update, don't stack. If a rule changes, edit the old one. Don't add a new one that contradicts it.
• Date the things that change. "As of October 2026, our price is..."
• Never store secrets. No passwords, no keys, no card numbers. Ever.
• Prune monthly. Delete what's no longer true.

The short version

• Memory is plain text files your AI reads first. That's all it is.
• Let it interview you. Then correct it and say "remember that."
• Keep one fact per file, update in place, and never store secrets.

If you'd rather have your memory set up with you, book a consulting call.

This is part 4 of the AI from 0 to 1 series. If you haven't set up the foundation yet, start with the free setup guide.

Questions people ask

What is AI memory?

A few plain text files your AI reads at the start of every session, so it already knows your business, your people and your rules.

Where does Claude or Codex read memory from?

Claude reads a file called CLAUDE.md and Codex reads AGENTS.md. The setup guide creates both with the same content.

What should never go in memory?

Secrets. No passwords, API keys or card numbers, ever. Store where a login lives, never the login itself.

Sources

1. Anthropic, How Claude remembers your project (Claude Code docs), 2026
2. Anthropic, Best practices for Claude Code, 2026