What lasts and what doesn't from one conversation to the next

This lesson is still being tested: it improves with learners' feedback. If something is unclear, it is the lesson that needs fixing, not you.

By the end of this lesson, you will be able to

say what your assistant keeps from one conversation to the next, and what it forgets;
see and correct what it has remembered about you;
write short instructions, applied to all your conversations;
open a conversation without memory, or outside your history.

Why it matters to you

Many people repeat the same instructions in every conversation (“use a casual tone”, “keep it short”), or on the contrary discover that the tool remembers things they would rather have kept to themselves. A few minutes of setup solve both: you write down once what should last, and you know where to look at what has been remembered.

Before you start


You need	Your assistant, free account
Cost	€0
Duration	35 minutes
Reminder	What you put in the memory and the instructions goes with all your conversations: nothing from the “never” list in module 1

Understand it in 3 minutes

A new conversation does not pick up everything that was said before. It benefits from what was kept: the memory (which draws on your past conversations, choosing what seems useful to it), your instructions, and the project you are working in.

What can last	Who writes it	What to know
The memory	The tool, as your conversations go along	It remembers what it judges useful, not everything. You can read it, correct it, pause it
The instructions	You	A few lines applied to all your conversations: tone, length, way of working
The projects	You	A space that groups conversations and documents on the same subject (with instructions, depending on the plan)
The history	Automatic	Your conversations stay in the list, so you can reread them; the memory can use them, without carrying over everything

What the official help pages say (read on 24/10/2026):

Claude: memory is turned on by default on the Free, Pro and Max plans. Claude remembers topics as your conversations go along, and you can tell it “remember this”. By default, it does not keep sensitive topics such as health or religion, unless you allow it. Searching your past conversations is reserved for paid plans. Projects are open to free accounts, 5 at most.
ChatGPT: the memory can remember preferences and details from your conversations, but not every detail. Your instructions apply to all your conversations; on the free plan, they are limited to 1,500 characters. Projects group conversations, files and instructions, depending on what your plan allows.

Warning: on Claude as on ChatGPT, deleting a conversation does not necessarily delete what the memory drew from it. To erase a piece of information, delete the memory entry in the memory settings, and the conversation where you gave it. Complete deletion on the provider’s side can take up to 30 days: Claude deletes a conversation from its servers within 30 days, and OpenAI may keep logs of a deleted memory for up to 30 days.

To remember: what really lasts is what you write yourself (your instructions, your notes, your request library). The memory is a useful extra, which you reread from time to time.

Do it together

The paths below come from the official help pages. The labels on your screen may be slightly different.

1Step 1: Ask what it knows about you

What have you remembered about me? Give me the list, without making anything up.

Read the list. If something is wrong, or too personal, you correct it in step 2. This answer is an overview, not a complete inventory: also read the memory page in the settings (step 2). On ChatGPT, even the settings summary may leave out details.

2Step 2: See and correct the memory

On Claude: click your initials at the bottom left, then the settings (“Settings”), “Your Claude” section, then “Manage” next to your memory. In the older layout: “Settings”, then “Memory”. Open a topic to read it, edit it or delete it.

On ChatGPT: “Settings”, then “Personalization”, then “Memory”. Depending on your account, you can correct the summary of what it remembers, or ask it not to mention a piece of information anymore.

Warning: asking it not to mention something anymore deletes nothing. According to OpenAI’s help page, to delete a memory you have to delete the memory entry and the conversation where you shared it (and, where relevant, the file or connected app that contains it).

Delete at least what is wrong. Then, depending on what you want:

Stop the memory: if you no longer want the tool to use its memory, you can pause it (Claude) or turn it off (ChatGPT) in the same place. Be careful, this does not erase what is already there: when paused, Claude keeps its existing memory, and turning off ChatGPT’s memory does not delete your past chats.
Erase everything: to erase, delete the memories one by one (on Claude, do it before pausing: the help page provides for this deletion when memory is on), or reset Claude’s memory (this is permanent); on ChatGPT, if the option is offered on your account, “Delete and turn off memory” deletes the information shown in the memory summary and turns it off, but not your past chats.

3Step 3: Write your instructions

On Claude: “Settings”, “Your Claude” section (or “Account” in the older layout), “Instructions” field. On ChatGPT: on the website or the desktop app, “Settings”, then “Personalization”; on iPhone or Android, “Settings”, then “Customize ChatGPT”. Check that customization is turned on (“Enable customization”), then fill in the “Custom Instructions” field.

Paste this template, then adapt it to your way of working:

Answer in English, in a casual, friendly tone.
Get straight to the point: the answer first, the details only if I ask for them.
Before a long piece of work, rephrase my request and ask me your questions.
When a piece of information may have changed, or you're not sure, say so, and suggest an official source where I can check.
If my idea has a flaw, tell me frankly.

You recognize the method: DISCUSS (rephrase), CHECK (the source), and the neutral question from lesson 2. On ChatGPT, the help page specifies that your instructions can also be used to improve the models, depending on your privacy setting (module 1, lesson 3).

4Step 4: Test in a new conversation

Open a new conversation and ask:

How do I prepare for a job interview?

Check: does the AI use a casual, friendly tone? Does it get straight to the point? If not, reread your instructions.

5Step 5: A conversation without memory, or outside your history

Outside your history, and without adding anything to the memory (the provider may still keep a copy, 30 days in general):

on Claude, an incognito chat: the ghost icon, at the top right, when you open a new conversation outside a project;
on ChatGPT, a temporary chat: open a new conversation, then choose “Temporary”. Before your first message, choose “Unpersonalized” if you don’t want it to use your memories or your instructions; this choice cannot be changed afterwards. It neither creates nor changes memories, and does not appear in the history, unless you save it.

Keeping the conversation, without using the memory:

on Claude: the + button, then turn off “Memory”, before your first message;
on ChatGPT: an “Unpersonalized” temporary chat uses neither the memory nor your instructions, but it is only kept if you save it.

What incognito does not do:

on Claude, an incognito chat is kept for 30 days by default, and your profile preferences still apply to it;
on ChatGPT, OpenAI may keep a copy of a temporary chat for up to 30 days, for safety;
A closed incognito chat cannot be reopened: copy what you want to keep first. And on ChatGPT, saving a temporary chat turns it into an ordinary conversation.

The “never” list still applies everywhere.

Your turn

Open your request library (resource below), “What my assistant keeps” section.
Note whether your memory is on, what you corrected or deleted, and where to open a conversation outside your history.
Copy your instructions into the library: if you change tools, you paste them back in one minute.
Use them for a few days, then change a line that doesn’t suit you.

Resources to download

My request library: your instructions, your settings, then your reusable requests (lesson 6).

Check your work

It worked if:

you can show, in your tool, where the memory and the instructions are;
you reread what the tool remembered about you, and deleted what was wrong or too personal;
your instructions fit in 10 lines at most, and a new conversation applies them;
you know how to open an incognito chat (Claude) or a temporary chat (ChatGPT).

Common mistakes:

believing the memory remembers everything: it chooses; what matters, write it in your instructions or your notes;
deleting a conversation thinking you are erasing the memory that comes from it;
long instructions that contradict each other: when they are short, they are easier to check and correct;
entrusting information from the “never” list to the memory or the instructions.

Keep it and pick it up later

Your instructions are written in your tool, and copied into your library.
Reread your memory once a month, or when an answer surprises you (“What have you remembered about me?”).

Going further

optional

On Claude, each project has its own memory. On ChatGPT, a project’s instructions take priority over your general instructions, inside that project.
In module 5, you will find the same idea again with an agent: a file of instructions that you write yourself, reread at the start of each session.

New words

5 words explained

Word	What it means
Memory	What the tool remembers from your conversations for the following ones, choosing what seems useful to it
Instructions	A few lines that you write, applied to all your conversations (“Custom Instructions” on ChatGPT)
Project	A space that groups conversations and documents on the same subject, with instructions depending on the plan
Incognito, temporary chat	A conversation that does not appear in your history and does not create memories, but may be kept for a while by the company
History	The list of your past conversations, which you can reread or delete

Still to verify: the exact on-screen labels of the settings (taken from the official English help pages, not from a live test). Claude’s help page contradicts itself on project instructions on the free plan (one page says they are reserved for paid plans, another says they are not): this lesson does not rely on it.