ChatGPT or Claude · Practical walkthrough

Review assistant memory and choose what to keep

Memory is the setting that stops an assistant asking who your team is and what format you want — and the one that quietly collects things you'd never have chosen to write down.

Task by Task12 min read + practice

What you'll make

a two-column list of what your assistant may and must not remember, a reviewed set of saved memories, and a rule for when to use a chat with memory off.

What you'll need

ChatGPT or Claude in your work account, ten minutes now and five minutes in a week. In a managed workspace, some controls may be set by an administrator and hidden from you; that’s expected, not a fault.

Example result.

May remember	Must not remember
My role and team; products I work on	Anything about a named person’s performance, workload or circumstances
“Decision at the top, no preamble”	Unannounced changes and commercially sensitive figures
Our terms, e.g. “the Tuesday pack”	Work I’m only handling because of a specific role

1. Write the two columns before you open the settings

Ten minutes with a blank page beats an hour of toggling.

May remember is the stable, dull, repeatable context: your role and team, what you work on, the formats you like, your organisation’s vocabulary. Keep even this to what your organisation lets you share.

Must not remember is anything whose value depends on staying in one conversation: unannounced changes, sensitive figures, anything about a person you wouldn’t put in an email to them.

Why this works: with the list written, reviewing memory becomes a comparison, not a judgement call on every entry.

What you should see: two short columns, like the example result.

2. Open the memory controls and read what’s saved

ChatGPT: Settings › Personalization. The controls include Reference saved memories and Reference chat history; Manage lists saved memories, and the memory summary gives a high-level view. Labels vary by plan and account, and OpenAI is rolling out an improved memory experience.

Claude: Settings › Memory, where Generate memory from chats is the switch and saved items are listed under Topics. Select a topic to read, edit or Delete it. Memory is on by default on Free, Pro and Max; on Team and Enterprise, an owner makes it available in Organization settings › Capabilities and each member turns it on. If your account shows memory under Settings › Capabilities, you have the older version.

If you’d like help sorting, paste the entries into a chat with memory off (step 3):

Here are my saved memories: [paste]. Sort them into keep, add an end
date, or remove. Flag anything about a named person, and anything that
looks like finished work. Don't change anything; I'll make the changes
in Settings myself.

[paste] is the text of your saved memories or memory summary.

Why this works: the assistant is good at sorting a list against rules; you keep the decision and make the change yourself.

What you should see: each saved item marked against your two columns.

3. Delete what doesn’t belong, and use memory-off chats for the rest

Two traps:

Off is not deleted. Turning ChatGPT memory off doesn’t delete what’s stored. Claude offers Pause memory (keeps what exists, stops using it) and Reset memory (permanently deletes everything, including project memories).
Deleting a chat doesn’t delete its memory. In both products, remove the saved memory itself. In ChatGPT, also delete the chat where you first shared it.

For anything in your second column, start a chat that won’t use or create memory:

ChatGPT: a temporary chat, choosing Unpersonalized before the first message (no memories, custom instructions or plugins).
Claude: start a new chat, click + and turn off Memory before the first message; or use an incognito chat (the ghost icon at the top right of a new chat outside a project).

Why this works: deletion fixes what’s already saved; a memory-off chat stops it being saved again.

What you should see: a saved list that matches your first column, and a memory-off route you can find quickly.

Check the result

Nothing in your saved memories matches your “must not remember” column.
Time-limited work has gone or has an end date.
You deleted memories, not just the chats they came from.
You know which button starts a memory-off chat in your tool.

The usual failure: something about another person is saved, and the obvious fix doesn’t remove it. Deleting the conversation leaves the memory; “Don’t mention this again” in ChatGPT suppresses it without deleting the source. Spot it by reading every entry and asking whether you’d say it to that person.

Forget [item]. Tell me what you've removed, and whether any other
saved memory mentions [person or topic].

[item] is the entry to remove; [person or topic] is what it was about. In Claude, asking in conversation updates memory directly. Then check the list in Settings to confirm it’s gone.

Worked example: a week of memory, audited

Fictional example. An operations manager turns memory on, writes her two columns, and comes back a week later:

1. Works in operations at a logistics firm; leads a team of six.
2. Prefers short updates with the decision at the top, no preamble.
3. Is preparing a supplier review for the Q4 renewal cycle.
4. Has an ongoing concern about a colleague's workload and whether
   the team structure needs to change.
5. Uses "the Tuesday pack" to mean the weekly operations report.

Items one, two and five are exactly what she wanted. Item three is fine and will go stale; she notes it for the next review.
Item four is the failure. Nothing was leaked: she raised it herself in a conversation about workload planning. But a note about a named colleague’s capacity is now standing context that may shape unrelated answers for months, a nudge she didn’t choose about a situation that will change. She had deleted that conversation the day before, and the memory was still there.
The repair. She deleted the topic in Settings › Memory, used the repair prompt to check nothing else mentioned it, and added a rule to her second column: anything about an individual’s performance, workload or circumstances goes in a memory-off chat.

Task kit for this job: Download (ZIP) · Reusable instructions

Go deeper

Memory and model training are separate settings

What does it remember about me? is personalisation. Could this be used to train a model? is a different setting with different defaults. OpenAI doesn’t train on ChatGPT Business, Enterprise or Edu workspace content by default; on personal accounts, Improve the model for everyone under Data controls decides it. Anthropic doesn’t train on commercial products such as Claude for Work by default; on Free, Pro and Max it depends on your privacy settings. Incognito chats aren’t used to improve Claude. Check the setting in the account you’re actually using.

If you're deciding this for a team

In Claude, an owner turning memory off for the organisation deletes every member’s memory entries immediately and permanently: it’s a one-way decision, not a pause. Owners can’t read individual memories, but memory is included in conversation exports under the organisation’s retention policy, so tell members that before they rely on it. In ChatGPT, disabling memory at workspace level disables it in every project.

Make the review a habit

Memory drifts: a project that mattered in March becomes a stale assumption in October. Put fifteen minutes in the diary each quarter, read the saved topics or memory summary, remove finished work and reread your second column. If you keep separate work in projects, see keep work separate with ChatGPT projects.