Module 5 · Protect the data
Data settings and memory in AI services
In short
The model does not know your account settings. Check them yourself: whether chats train the model, how long they are kept and what memory holds. Use a work account for work, and clean out memory regularly.
After this lesson you can
- Check an AI service's data settings: whether chats are used for training, and how long they are kept
- Review and delete the contents of memory regularly
- Tell a personal account apart from a work account your organisation manages
Terms in this lesson
Understand
Almost every AI service has a data settings page, and almost nobody opens it. Yet three settings there decide what happens to everything you have ever typed.
1. Used for training or not. Model training means the provider uses large collections of text to build or improve a model. On many services for individuals, your chats may be used to train the next model unless you turn it off. There is usually one switch with a name along the lines of "use my data to improve the model". If it is on, anything you type may become training material.
2. How long it is kept. Chat history is kept so you can go back to it. Deleting a chat from the list does not always remove it from the server at once; many providers keep copies for a while for safety or legal reasons. The provider's privacy policy usually states how long. Many services also have a temporary chat mode that stays out of history and is not used for training, useful for questions you do not need to keep.
3. Memory. Memory is the feature that makes the app remember things about you from one conversation to the next: your role, your style, your projects. It differs from history. History is an archive of old chats; memory is a short note the model reads in every new chat. Memory saves time, but whatever is in it travels everywhere, including things that are sensitive or no longer true.
One thing that often surprises people: the model itself does not know your account settings. If you ask "is this chat used for training?", the answer is only a confident-sounding guess. The real answer is on the settings page and in the privacy policy, not in the chat box.
Providers often treat work and personal accounts very differently. Plans for organisations usually come with a contract that governs data use, and at many providers work-account data is not used for training by default. But "usually" is not a guarantee: read the terms, or ask the IT team that manages the subscription.
The most common problem is not a wrong setting but the wrong account. Someone uses a free personal account on their phone to tidy an office document, because it is quicker. The office data then sits under the personal account's terms, outside the organisation's control, recorded nowhere. The rule is simple: office work in the approved work account; the personal account for personal matters.
For long pieces of work, do not use the app's memory as the place to keep progress. Memory is brief, can change, and is not fully under your control. Important progress goes in a notes file you own; how to do that is in the lesson on projects and standing instructions.
A step specific to OpenAI: check the latest version.
A step specific to Anthropic: check the latest version.
Before
The prompt
Are my chats here used to train the model? I want to paste customer data.
What came back
No, your conversations are safe and are not used for training. Go ahead and paste the data.
After
The prompt
I need to check whether my account uses chats for model training and how long chats are kept. I know you cannot see my account settings. Below I paste the relevant part of this provider's privacy policy. [privacy policy text] Show me the exact sentences from that text that answer my two questions. If the text does not answer them, say so. Then list three things I must check myself on the settings page.
What came back
Relevant sentence: "We may use your content to improve our services unless you turn this off in settings." On retention, the text you pasted does not state a period. What you need to check yourself: 1. Whether the training-use switch is on. 2. Whether memory is on and what it holds. 3. Whether this is a work account under an organisation contract or a personal account.
What changed
- The first answer sounds sure but has no basis: the model cannot see your account settings.
- Three things change the result: the policy is pasted as material, exact quotes are requested, and "not stated" is allowed as an answer.
- The final decision still rests on the settings page, which you open yourself.
Common mistake
- Mistake
- Letting memory store anything for months without ever opening it.
- Why it hurts
- Memory may hold client names, health details you mentioned in passing, or an old role that no longer applies. All of it is read in every new chat, can surface in answers for unrelated work, and can skew answers without you seeing why.
- Do this instead
- Schedule a memory review once a month. Delete what is sensitive, wrong or irrelevant. Turn memory off for an account used for many clients at once.
Try
This is an HR officer's memory. Keep or delete? Guess, then reveal the reason.
Keep. Useful and not sensitive.
Delete. Third-party personal data that would travel into every chat.
Delete. Your own health data does not belong in a work tool's memory.
Delete or update. Stale memory skews answers.
Take home
Monthly check
- The training setting is still as I chose
- I read memory and deleted what is sensitive, wrong or stale
- I deleted old chats with work data that I no longer need
- Office work happens only in the work account
- Progress on long work is in my own file, not in memory
What went wrong for us
The ledger file that saved a programme
- Context
- An agent work programme with more than a dozen tasks ran for days across several sessions.
- What happened
- Mid-task, the session stopped on a usage limit. The old conversation could not be continued and its context was gone.
- Impact
- Almost none. A new session picked up at the right task within minutes.
- Root cause
- An agent's memory only lasts one session. Without notes outside the session, the work has to be guessed again from git history.
- What we did
- The new agent read the progress ledger, saw the last reviewed task, and carried on from there.
- The guard we keep now
- Every long programme has a ledger file: tasks, status, commits and owner decisions. Agents write to it every time a step finishes.
The lessonIf work outlasts one session, its memory has to be written outside the session.
What it means for your workFor work longer than one chat, write progress notes in your own file each time a step is done.