Persistent Context: Projects, Memory & Custom Instructions
The four layers of context, how to write instructions and projects that work, and how to manage memory and privacy.
Tired of explaining yourself in every new chat? AI tools can remember things for you. Put facts about you in a short profile, put project details in a project, and keep your own copy in a notes file. Never save passwords or private client data.
- Custom instructions
- A short note about you that the AI reads every time.
- Memory
- Facts the AI saves from your chats. Check it now and then.
- Project
- A folder with instructions and files for one ongoing piece of work.
Explaining your job and your preferences in every new chat is the biggest time waste in everyday AI use, and you can avoid it. Every big assistant now has ways to remember things for you. The names and details differ, and they change often. So do not memorize product features. Learn the four layers of context that all of them offer in some form, and learn where each kind of information belongs.
The four layers of context
- Global instructions: a short profile that applies to every chat. It says who you are, how you like answers and your standing rules. It rarely changes.
- Memory: facts the assistant saves about you and your work over time. It saves them by itself or when you ask. It is useful but hard to see into, so review it.
- Project or workspace: a container for one body of work, with its own instructions and files. Every chat inside it starts already briefed.
- In-chat context: what you paste or type right now. It is the most specific and it lasts the shortest time.
The rule: put each piece of information at the broadest layer where it is always true and useful. Your writing style goes in the global instructions. We are writing the Q4 investor update, and here is last quarter's version goes in a project. Today's numbers go in the chat.
Writing global instructions that work
About me: I'm a marketing manager at a 40-person B2B software company. My readers are busy non-technical executives. I'm not a native English speaker; plain, simple wording helps. How I like answers: - Lead with the answer or recommendation, then the reasoning. - Short paragraphs. Use bullet points only for lists of parallel items. - Don't flatter me or pad. If my idea has a problem, say so directly. - If you're unsure or the information may be out of date, say so. - If my request is ambiguous and the answer would differ a lot, ask one clarifying question first. Defaults: British spelling. No emojis. Don't use the words "delve", "leverage" or "unlock".
Building a project that remembers
A good project has three parts: instructions, reference files and a clear scope. The instructions say what the project is, who the audience is, the tone and format rules, and what a finished piece looks like. The files are the source of truth the AI should rely on, such as brand guidelines, your best past examples, product facts and a glossary. The scope keeps it focused. Make one project for each ongoing body of work, not one giant project for everything.
PROJECT: Monthly customer newsletter GOAL: Draft each issue from my rough notes in our brand voice. AUDIENCE: Existing customers, mostly operations managers; they skim. SOURCES OF TRUTH: Use only the attached brand-voice guide and product-facts sheet for claims about our product. If a claim isn't in them, flag it as "needs verification". FORMAT: Subject line options (5), 80-word intro, 3 sections with bold headings, 1 link roundup, a closing question. Under 450 words total. PROCESS: First restate what you understood from my notes and list gaps. Wait for my answer before drafting if more than two facts are missing. DON'T: invent customer quotes, statistics or features.
Managing memory sensibly
- Read what has been saved about you. It is often incomplete, out of date or slightly wrong. Wrong memory quietly bends every answer.
- Tell the assistant clearly what to remember, for example remember that I prefer short answers, and what to forget.
- Use a temporary or incognito chat for anything sensitive or one-off, so it does not enter memory.
- Check it again every few months. Roles change, projects end and preferences shift.
- Never rely on memory for facts that must be right. Put those in project files you control.
Privacy: what not to store
Persistent context is handy because it keeps your information. So keep some things out of memory and global instructions. These are passwords, bank details, health details, other people's personal data and confidential client or company material. Check your workplace's AI policy and your plan's data settings. Some plans use your chats to improve their models unless you opt out. Company accounts may have different rules.
Own your context
Memory features belong to one company and do not move with you. Keep a master context document in your own notes app. It holds your profile, your standing preferences, your project briefs and your best prompts. When you switch tools or an account resets, you paste it back in minutes. Treat the assistant's memory as a handy copy of your document. Your document is the original.
- ✓There are four layers: global instructions, memory, projects and the chat. Put information at the broadest layer where it is always true.
- ✓A project is instructions, source files and a clear scope.
- ✓Review and tidy your memory, and keep sensitive data out of it.
- ✓Keep your own master context document, so you are not tied to one tool.
Set up your context in 30 minutes
- Write your global instructions, using the example as a model: about you, how you like answers, and your defaults. Keep them under 200 words.
- Make a project for one job you do again and again. Add instructions, the two or three files it should treat as true, and a clear definition of done.
- Open a new chat inside the project and make a one-line request. Compare it with how you would have had to brief the AI before.
- Save all of this in one note called My AI context, which you own.
Q1Where should we are drafting the Q4 investor update live?show answer
In a project. It is true for a whole body of work, not for every chat, and it helps to attach files such as last quarter's update.
Q2Why review the assistant's saved memory from time to time?show answer
Saved facts can be wrong or out of date, and they quietly bend every answer. Tidying them keeps the assistant's picture of you accurate.
Q3Why keep a master context document outside the tool?show answer
Memory and projects do not move between companies' tools. Your own document lets you rebuild your context quickly after you switch or after a reset.