ChatGPT memory and personalization without losing control

ChatGPT can tailor replies using custom instructions (Customize ChatGPT), saved memory, reference to chat history and other context depending on your plan and region. That saves repeating the brief, but it can also retain details you did not mean to keep. The official memory guide is at https://help.openai.com/en/articles/8590148-memory-in-chatgpt; custom instructions are explained at https://help.openai.com/en/articles/8096356-custom-instructions-for-chatgpt. In about 18 minutes you get a repeatable flow: what belongs in each layer, how to review what is remembered and when to chat without personalization.

1. Steps 1

   Decide what goes in instructions vs memory

   Before touching settings, write two lists in a notes app. List A — stable instructions for most chats: role, tone, format, bans (same rigor as chatgpt-primeros-pasos). List B — facts that change or are sensitive: allergies, client names, budgets, health details. Custom instructions are direct guidance you define; memory can pick up preferences and details from conversations and other sources available to your account. Do not put secrets or PHI on either list if your workspace is not meant for that. If the work lives in a container with fixed files, consider chatgpt-projects-primer-flujo in parallel: project instructions do not replace auditing global memory.

   What you should see: Two short lists: A for Customize ChatGPT, B marked «only if needed and reviewable».

2. Steps 2

   Open Customize ChatGPT and write custom instructions

   On https://chatgpt.com, open Settings (profile icon or menu) → Personalization → Customize ChatGPT. Fill what ChatGPT should know to work with you and how you want answers: default language, detail level, preferred format (bullets, tables) and rules like «mark [MISSING] if there is no data». Keep the block concise; a dense internal policy paragraph fits better in a Project or Custom GPT (chatgpt-gpts-cuando). Save changes. These instructions are separate from memory: text you control explicitly.

   What you should see: Customize ChatGPT screen with both fields filled and saved.

3. Steps 3

   Review Memory controls in Settings

   Go to Settings → Personalization → Memory and find the toggles your account shows. You will usually see a main Memory switch and, depending on the experience, options such as Reference saved memories, Reference chat history, Memory summary or Manage. Turn Memory on only if you want persistent personalization; turning Memory off does not delete past chats by itself, but ChatGPT stops creating new remembered information until you turn it back on. If you have Reference chat history, know that turning it off schedules deletion of information derived from history on OpenAI systems over the timeframe their help article describes (up to about 30 days), without removing chats from your list. Exact labels can vary — trust what you see, not old screenshots.

   What you should see: Memory section with at least one toggle and a Manage or Saved memories link if applicable.

4. Steps 4

   Inspect the summary and saved memories

   Click Manage, Memory summary or Saved memories (depending on your version) and read what ChatGPT may use to personalize. You can ask in chat: «What do you remember about me?» to cross-check. Correct in the text field, select text and add a correction, or use Don't mention this again when shown to reduce future references without deleting the original source. To remove an item from the summary, use Delete or the (•••) menu options. Saved memories are explicit entries stored apart from chat history: deleting the chat where one was created does not always delete a separate saved memory — check both.

   What you should see: A memory list or summary with at least one recognizable item; edit or delete available.

5. Steps 5

   Use Sources and ask to forget when needed

   On personalized replies, check whether Sources appear below the message: they may cite a past chat, custom instruction, saved memory or file. Open them and mark relevance or remove the source if you do not want it to influence answers. Say in plain language: «That is wrong, update it» or «Do not use this in the future». For a saved memory, «Forget that…» or delete it in Memory settings. If you share the chat by link, memory Sources are not included on the shared link per OpenAI’s documentation. After deletion, updates can take time to propagate; do not assume instant effect on every device.

   What you should see: A reply with expandable Sources and, after a correction, fewer references to a wrong fact in new chats.

6. Steps 6

   Temporary chat when you do not want memory or instructions

   For sensitive, one-off topics that should not feed personalization, open Temporary chat from ChatGPT’s menu. Before you start, choose whether to use existing memories and instructions or Unpersonalized / without personalization — you cannot change that choice after the first message. A temporary chat does not create or update memories even if Memory is on globally. If you save a temporary chat, it becomes a regular chat and your personalization settings can apply; with Memory on, ChatGPT may use that conversation for future replies. See OpenAI’s temporary chat help linked from the memory article for details.

   What you should see: Temporary chat picker with personalized vs Unpersonalized before the first message.

7. Steps 7

   Scope memory inside a Project (when available)

   If you use ChatGPT Projects (chatgpt-projects-primer-flujo), check Project settings for project-only memory when your workspace offers it (regulated or team environments). That mode limits which conversations can reference each other: chats inside the project should not pull from threads outside, and vice versa, per OpenAI’s improved memory documentation. Create one project per client or workstream and keep long threads there. Project instructions add container context; they do not replace auditing global memory if you mixed loose chats on the same topic. On plans without project-only memory, still separate projects by case and clean global memory when you change context.

   What you should see: Project settings with project-scoped memory option, or at least project instructions visible.

8. Steps 8

   Privacy checklist: what not to save

   Close the flow with written rules in your notes: do not ask to remember passwords, tokens, card numbers, full government IDs, detailed medical diagnoses or non-public trade secrets. On personal accounts, review Data controls if you want to limit chat use for model improvement (Improve the model for everyone); on Business, Enterprise or Edu, OpenAI states workspace content is not used to train models by default. If something sensitive already entered memory, delete it in Manage, remove the source chat and drop files in Library or connected apps that contain it. For work that only needs format, stick to custom instructions without identifiable facts.

   What you should see: A five-item «never in memory» checklist and memory reviewed without critical data.