What Claude remembers on its own
Claude has a built-in memory. On Free, Pro and Max plans it's on by default; on Team and Enterprise plans, an owner turns it on and each person opts in. You can open Settings › Memory to read, edit or delete what it has saved, pause it, or reset it entirely.
A few useful details: each Project keeps its own separate memory, incognito chats are never saved, and Claude avoids saving sensitive topics and things like ID or account numbers unless you allow it.
Where built-in memory stops
- Deleting a chat doesn't delete what was learned from it. Memories made from a conversation stay until you remove them in Settings.
- Moving it is copy-and-paste. Anthropic offers import and export, but it's manual: you ask Claude to write out its memories and paste them elsewhere, and Anthropic describes import as "experimental and still in development".
- Some organisations can't use it. It isn't available to organisations with HIPAA, public-sector or custom data-retention agreements.
- It's Claude's. It lives inside one company's product. Your ChatGPT or next year's AI can't read it.
Built-in memory helps Claude remember you. A memory you own helps every AI know your business.
Built-in memory vs a memory you own
| Claude's memory | Your own memory (in Notion) |
|---|
| Who can read it | Claude | Claude, ChatGPT, Grok, your team, you |
| What's in it | What Claude decided to save | What you decide: people, rules, projects, routines |
| Moving to another AI | Manual copy and paste | Point the new AI at the same pages |
| Your team | Each person's memory is separate | One shared memory, with permissions |
They aren't rivals. Keep Claude's memory on for small personal preferences; keep the business itself (your clients, your promises, how you work) in pages you own.
Set up your own in Claude
Claude can read Notion through a connector. Claude's connectors work on web, desktop and mobile, and Anthropic's own list includes Gmail, Google Calendar, Google Drive, Microsoft 365 and Slack, plus custom connectors on paid plans. The setup is three steps: give your memory somewhere to live (a Notion workspace), teach Claude where it is (project instructions), and connect your email and calendar so it sees your week.
What it costs:
Claude Pro is $20 a month billed monthly, or $17 a month billed yearly. Notion's free plan is enough to start.
Why most AI assistants don't stick
Most people have tried an AI assistant, and it fizzled. Usually the tool was fine. It just never learned what your days actually hold: who matters, what comes round every week, which hours are off-limits, how you like to hear bad news.
You wouldn't hand a new hire a laptop and expect them to know your clients by lunch. You'd spend their first week showing them how things work. An AI assistant needs the same week.
The shape of week one
- Ten areas, six that matter first. The first six cover the essentials. The other four (health, home, hobbies, interests) fill in over time.
- About ten minutes a day. In whatever pieces suit you. Nobody finishes in one sitting, and nobody should try.
- A clear finish line. When the six are done, your assistant says so once, and then stops treating onboarding as homework.
The six essentials, and what each one asks
1. What fills your days. Your work in your own words: a business, a job, a household, caring for someone, or a mix. And one plain question, asked once: is any of it a business you run?
"What do you personally decide that nobody else can?"
2. What comes round, and where you want help. The things that repeat every week, month or season, what "done well" looks like for each, and what kind of help you'd like: a reminder, a draft, a plan, a list.
"For the ones that matter, what's the sentence you'd say if it went right?"
3. Your calendar rules. A handful of rules it can check before it books anything. "No meetings before ten on Tuesdays" is a rule. "Protect focus time" isn't.
"When a request breaks one of these rules, do you want me to hold the line and tell you, or book it and flag it?"
4. What's live right now. Everything in flight, with an owner and a date. About half won't have either, and seeing that is the point.
"Who is waiting on something from you right now?"
5. Your people. Who works with you, who you deal with most, who at home has a claim on your calendar, and who you talk to completely differently from everyone else.
"Is there anyone you communicate with completely differently from everyone else, and why?"
6. How you like to be told things. The update you liked, the one that annoyed you, and how you want hard news.
"Do you want bad news the moment someone knows, or packaged with a plan? Which one have you actually complained about before?"
Habits that make it work
- One question at a time. A numbered list gets three answers and an abandoned form.
- Ask about last week, not about who you are. Nobody can answer "what do you value?" Everybody can answer "what was wrong with the last message that annoyed you?"
- Read before asking. If your email and calendar are connected, a good assistant offers to read them and draft the answers for you to correct. Memory is the worst source there is. Your sent mail holds the promises you've already forgotten.
- A number, a name or a date beats an adjective. "We deliver frequently" isn't an answer. "Tuesday and Friday" is.
- Never invent. If you didn't say it, it doesn't get written down. A skipped question is marked as skipped, not left looking blank.
- Come up for air. Every few questions, it says what it wrote down and checks you're fine to keep going.
- Your work always wins. Onboarding never gates help. If you need an email drafted mid-interview, it drafts the email and picks the thread back up later. An assistant that makes you answer questions before it helps is a bulldog, not an assistant.
After week one: it learns from the work
Once the six are done, the questions stop. From then on, most of what your assistant learns arrives while it works: a calendar rule it works out while booking something, a person it meets while drafting to them, a preference you mention in passing. It writes each one down where it belongs and tells you in a line when it helps: "I've noted that Fridays are off-limits."
The later areas aren't a to-do list. They come up only when a moment calls for them.
Start yours today
Starter prompt (paste it into any AI, then answer one question at a time):
I want to onboard you as my assistant, the way I'd onboard a new hire.
Interview me, one question at a time, about six areas in this order:
1. What fills my days, and whether any of it is a business I run.
2. What comes round every week or month, what "done well" looks like, and where I want help.
3. My calendar rules: checkable rules you can test a request against.
4. What's live right now, with an owner and a date for each.
5. My people, at work and at home.
6. How I like to be told things, including bad news.
Ask about real recent examples, not abstractions. Never invent an answer.
If I skip something, write "skipped".
Keep a running document of my answers, one section per area, and show it to me
at the end of each sitting. We'll do about ten minutes a day.
Keep what it learns. Without a memory of its own, your AI forgets this by next week. Copy the document into your AI's project instructions at the end of each sitting, so every chat starts from it.
With The Modern Admin
Every assistant we set up runs this onboarding for you. The interview lives inside your assistant, so you just talk. Each answer is written straight into your workspace in Notion, on pages you can read and correct. It knows where you left off, so it picks the thread back up in a line ("Last time we got through your work. Next is your people, ten minutes or so"), and it never asks the same thing twice.
