Context is just familiarity you build on purpose.
You wouldn't expect a stranger to know what you mean. AI works the same way. There are three levels of familiarity, an acquaintance, a work friend, and the friend who really knows you, and both ChatGPT and Claude build them the same way. This page maps exactly where each level lives, what to put there, and how long it sticks.
- The three levels explained in plain language: acquaintance, work friend, the friend who really knows you
- Exactly where each level lives in ChatGPT and in Claude, side by side
- What to put in each level and how long it actually lasts
- The one prompt that documents what the AI already learned, so you paste it forward instead of retyping it
The three memories, side by side
Same three levels, ChatGPT and Claude call them slightly different things and put them in slightly different menus. Each card below is copyable if you want to save it as your own reference notes. Verified against current ChatGPT and Claude documentation and help articles, accessed Aug 9, 2026. Menu names move over time, so if a label below doesn't match what's on your screen, trust what's on your screen and use this as the map, not gospel.
Level 1: Chat, one conversation
CHATGPT Where it lives: the conversation you're in right now. What to put there: nothing to set up, just talk. Anything you haven't confirmed into memory only exists in this thread. How long it lasts: until you close the chat. A brand new chat starts cold unless account-wide memory (Level 3) carries something forward. CLAUDE Where it lives: the conversation you're in right now. What to put there: nothing to set up, just talk. How long it lasts: until you close the chat, unless Claude's memory (Level 3) has already picked up a summary of it.
This is the level almost everyone is stuck on. If you keep re-explaining yourself from scratch, you're living here.
Level 2: Project, one specific thing
CHATGPT Where it lives: open a Project, then its project instructions and files. What to put there: instructions and files for that one project. Project instructions take precedence over your account-wide Custom Instructions while you're inside it. How long it lasts: every chat inside that Project, until you edit it. A Project can also run in project-only memory mode, so it stops pulling in your account-wide saved memories. CLAUDE Where it lives: the Project's own settings, project instructions plus Project knowledge (the files you attach). What to put there: instructions and reference files for that one project. How long it lasts: every chat inside that Project. Each Project also keeps its own separate memory space and summary, so it doesn't mix with your account-wide memory or other Projects.
This is the work-friend level. Common ground already exists, but only inside that one Project.
Level 3: Settings, everywhere
CHATGPT Where it lives: Settings, then Personalization. Custom Instructions is the explicit box you write yourself. Memory (saved memories plus reference chat history) is automatic, review it under Manage Memories. What to put there: Custom Instructions: the handful of things true about you everywhere. Memory: nothing to write, it saves facts as you go, you edit or delete what it got wrong. How long it lasts: every chat, every Project, until you edit or turn it off. It's not unlimited, check in-app for the current limit before assuming it holds everything. CLAUDE Where it lives: Settings, then Memory (or Settings, then Capabilities, on the legacy or Enterprise experience). What to put there: nothing to write by default, Claude generates memory entries from your chats, your role, working style, and preferences. Review or edit what it's captured. How long it lasts: account-wide by default, kept per Claude's normal chat data retention. Step inside a Project and it switches to that Project's own memory space instead.
This is the friend-who-really-knows-you level. It follows you everywhere, so keep it to what's actually true everywhere.
The move: ask it to write down what it learned
Everything we've covered in this chat, turn it into a short set of instructions for how to help me. Write it so I can paste it into a Project's instructions or my account's Custom Instructions. Keep it to what's actually true about me, not a transcript of this conversation.
Run this at the end of any chat that taught the AI something worth keeping. Paste the result into Level 2 if it's specific to one thing you're working on, or Level 3 if it's true everywhere. Want the longer version, fifteen questions instead of one chat? The Interview Prompt below does that from scratch.
The friendship version
Acquaintance = a regular chat
Might pick up a few things mid-conversation. Every new chat needs some catch-up. You can't say "you know what I mean" and expect it to land.
Work friend = a Project
Common ground already exists, but only for that one thing. Every chat inside the Project knows it. Nothing outside the Project does.
The friend who really knows you = account-wide memory and Custom Instructions
How you communicate, what you care about, everywhere, every chat, every Project. Just like a real friendship, if you change, you have to say so.
The ChatGPT Interview Prompt
Fifteen questions, one at a time, then it writes its own instructions and saves them to memory. The long version of the meta-prompt above.
Read more →Prompt Starter Pack
25 copy-and-paste prompts and the CLEAR framework, for once your AI actually knows who it's talking to.
Read more →This map is free. The paid packs go deeper: full prompt libraries, role-specific versions, and workflow chains you can run start to finish. One-time purchase, instant download.
See the paid packsReal builds, no jargon, no gatekeeping. One issue a week.
Anyone can learn this. The gate is intimidation.
@keybuilds.ai →