Sam Davis · Sydney · October 2026Most Claude or ChatGPT chats start close to zero, because the context you give them lives inside one chat, and chats end. Write what it needs to know into files it reads every time, and your AI use, and your team’s, starts building on itself.
Each new chat is a new hire on their first morning
It’s Monday. You open a new chat to draft something for a client, and for the third time this month you explain who you are, what you sell, who this client is, and how you like things written. Twenty minutes later you have a decent draft. Next week, the same speech.
Most people end up with a pile of named chats they return to: one for the newsletter, one for proposals, one for that big project. It helps for a while. Then the chats get long, or you start a fresh one, and you’re back to the introductions.
In a team it’s worse. Three people using AI, each with their own history. Someone finds a prompt that works, and it lives in their account, where nobody else will see it.
The model only knows what’s put in front of it
A language model doesn’t carry anything from one conversation to the next unless something puts it back in front of it.
Both Claude and ChatGPT have added memory features that carry some of this between chats, and they help. What they remember sits inside one person’s account, though, so a colleague’s chats don’t benefit. The fix that holds up is to write it down somewhere the AI reads every time.
You can start with one page and a Project
- Get interviewed for your context page. Open a chat and paste: “Interview me about my business so you can write a one-page brief another assistant could work from. Ask one question at a time and stop when you have enough.” Let it ask about what you do, who you do it for, your offers, the people who matter and how you write. Save the page it writes.
- Make a Project for each area of work. In Claude (ChatGPT has the same idea), a Project holds files and instructions that every chat inside it sees. Put the context page in each one, plus the documents that area needs.
- End good chats with one line. “Write down anything from this chat that should carry into the next one.” Add what it writes to the Project’s files. Then the next chat starts with what the last one learnt.
- Save prompts you’ve used twice. Keep them in a shared document the team can copy from. If your team is on one of Claude’s team plans, Projects can be shared with colleagues too.
Give it a few weeks and the chats tend to start with less explaining and end with better drafts.
The manual version breaks at the copy and paste
You’re the one carrying memory between chats. Files drift out of date unless you update them, so the AI knows your business as it was when you last did. And nothing happens while you’re away from the keyboard: no morning summary, no nudge that a client has gone quiet.
An assistant with its own memory does the carrying
Crads-AI is one way to hand that job over. It’s an AI assistant, open source and built on Claude, with a memory of its own. It keeps everything it knows about you as plain pages in a “brain”, backed up to a private GitHub repository in your own account. It checks those pages before it answers. An hour-long interview writes the first version; after that, a capture at the end of a session writes down what’s new, and a prompt you keep reusing can become a skill you run by name.
For a team, there’s a company brain every person’s assistant reads first. Facts from each person’s work are published to it with a line saying who they came from, and changes to policies or prices wait for the owner’s approval.
The thinking happens at Anthropic, on your own Claude account, so what you type and the pages it reads go there. The brain sits on a server in your own hosting account or on your own computer (on your own computer, nothing runs on a schedule). Crads-AI holds none of it.
One call can show what your AI should know first
The steps above work on their own. If you’d like help deciding what to write down first, or how a team should share it, book an AI problem-solving roadmap: a 30-minute call about how you and your team use AI now, a written report on how to make that use add up (the do-it-yourself route included), and a second 30-minute call to go through it together.