ChatGPT memory and custom instructions at work: a practical setup guide

Know which kind of context you are changing

Custom instructions are guidance you write deliberately: how ChatGPT should respond and what it should know about your preferences. OpenAI says they apply immediately to chats and can be edited or removed for future responses. Memory is different. When enabled, it can draw on saved memories, past chats and other available sources, depending on your account. It does not retain every detail or consult every past conversation for every answer.

For work, put a stable preference in custom instructions, such as Australian English or a preference for short decision summaries. Put today's client facts, figures and deadline in today's chat or an approved project. A memory is a convenience for recurring context, not an authoritative company record. If a policy changes, update the policy at its source and check any saved instructions that mention it.

Write a small set of useful custom instructions

Open Settings, then Personalization or Customize ChatGPT, and find Custom Instructions or the Enable customization control shown on your device. OpenAI lists the feature across its plans, but the exact controls can vary by platform and workspace. Save guidance that will still make sense next month. Avoid copying a whole job description into a global field when only one project needs it.

Try three rules: your preferred language, the structure of a useful answer, and how to handle missing facts. These are instructions about the output, not permission to invent facts or act on your behalf.

1. Write two or three short rules based on work you actually repeat.
2. Save them in Custom Instructions and open a low-stakes chat with fictional information.
3. Ask for an update that includes a known gap. Check whether ChatGPT marks the gap instead of filling it in.
4. Edit one rule if the answer is too long, too certain or in the wrong format. Test again before adding more.

Use Australian English. For work updates, start with the decision needed, then give up to three supporting points. Separate confirmed facts from assumptions. If an owner, date or figure is missing, say it is not recorded; ask me before inventing one. Do not claim you checked a source unless you actually opened it.

Review memory instead of assuming it is correct

If Memory is available, inspect Settings > Personalization > Memory. OpenAI says controls can differ by plan, region, platform and workspace. Depending on your account, you may see saved memories, reference chat history, a memory summary, Manage or Sources beneath an answer. You can also ask ChatGPT what it remembers, but check the actual controls and sources rather than treating that reply as a complete audit.

An old preference can quietly steer a new answer. For example, you may once have asked for sales language that no longer fits a compliance update. Look for stale roles, names, locations and output preferences. Correct or remove individual items where those controls exist. OpenAI notes that deleting a chat alone does not necessarily delete a separate saved memory, and switching Memory off does not delete past chats. Review the relevant source as well as the remembered item.

Choose a temporary chat deliberately

A temporary chat stays out of normal chat history and does not create or update memories while it remains temporary. It can still use existing memories and custom instructions if you choose Personalized before the first message. Choose Unpersonalized if you want it to ignore those settings. OpenAI says that choice cannot be changed after the conversation starts, and saving a temporary chat converts it to a regular chat with your account's usual settings.

Temporary does not mean it is safe to paste anything. OpenAI says a copy may be kept for up to 30 days for safety, and workplace restrictions still apply. A customer's contract, employee record or unreleased financial result should only be used in an approved account and workflow. For a practice exercise, make up the facts.

Check privacy and workplace controls separately

Memory, chat history and model-improvement controls answer different questions. On a personal account, OpenAI says Improve the model for everyone is managed in Data controls. Turning it off affects new conversations used for model improvement; it does not delete existing chats. OpenAI says Business workspace data is not used for model training by default, and workspace owners can control whether Memory is available. Ask your workplace which account and materials are approved before you rely on a personal setting as a company policy.

Do not put passwords or sensitive personal details into a reusable instruction or memory merely to save typing. Store approved templates where your team already maintains them. When a response cites an old conversation, file or connected app, check that source and its date before you carry the statement into a document.

Use a small test and maintain it

Test the setup with a fictional weekly update: one completed task, one unresolved delay and one missing owner. A useful answer should follow your requested structure, keep the delay visible and leave the owner unassigned. If it invents an owner, correct the result and tighten the instruction. Do the same test after a major change to settings or when an answer suddenly sounds unlike your normal work.

This setup should remove repeated briefing, not create a hidden second policy book. Keep the global instructions short, inspect Memory occasionally, and put facts in the current task. For a team exercise built around your actual workflow, use the workshop enquiry on the Clearer With AI homepage.

Common questions

Are custom instructions the same as saved memories?

No. You write custom instructions directly. Saved memories are separate context ChatGPT can use when Memory is enabled. Check and manage each in its own setting.

Does Temporary Chat ignore everything ChatGPT knows about me?

Not automatically. OpenAI now lets you choose Personalized or Unpersonalized before a temporary chat begins. Choose Unpersonalized to avoid using existing memories and custom instructions; neither mode creates new memories while it stays temporary.

Will deleting an old chat remove a saved memory from it?

Not necessarily. OpenAI says a separately saved memory can remain after the original chat is deleted. Review and remove the saved memory and its source where appropriate.