In June 2026, OpenAI changed how ChatGPT remembers. The newer memory system, now running on both paid and free accounts, no longer depends on a short list of facts you manually save. It reads across your past conversations in the background and builds a running picture of who you are, what you work on, and how you like your answers framed. Paid accounts also received roughly twice the previous memory capacity. For one person, that means less repeating yourself. For a business, it means your assistant is quietly accumulating a record of your operations, your clients, and your commercial decisions.
That record is helpful right up until it becomes a liability. If your team pastes client details, contract terms, or financial figures into ChatGPT, some of that can now persist across sessions without anyone choosing to keep it. This guide covers what the memory feature holds, what the Privacy Act expects of you, and the exact settings that put you back in control.
What ChatGPT memory actually stores now
Memory works in two layers. The first is the set of details you or the assistant explicitly record, which you have always been able to open and edit. The second, newer layer is the picture ChatGPT forms on its own by reviewing earlier chats. This second layer is the one that catches people out, because nobody pressed save. It can include the kind of work you do, the tools and people you mention often, your recurring projects, and the writing style you prefer.
OpenAI added a readable summary page that shows what the assistant has concluded about you, with controls to add, edit, or remove any single item. That visibility is the practical starting point for any business that wants to manage its exposure rather than guess at it.
Why this matters under the Australian Privacy Act
If your business has an annual turnover above $3 million, or it handles health information, you are almost certainly covered by the Privacy Act 1988. The Australian Privacy Principles set out how you may collect, use, store, and disclose personal information. Two principles are directly relevant here. APP 6 limits using personal information for a purpose the person did not expect. APP 11 requires reasonable steps to protect that information from misuse and unauthorised access.
When a staff member pastes a customer's personal details into ChatGPT and the memory system retains them, you have arguably disclosed that information to a third party and created a copy you no longer directly hold. If those details travel to overseas servers, APP 8 on cross-border disclosure also applies. None of this makes ChatGPT off limits for Australian businesses. It does mean the memory feature needs a deliberate decision, not a default you never reviewed.
Client personal information: names, contact details, and case notes that identify a specific person.
Health and other sensitive information, which carries stricter handling rules under the Privacy Act.
Financial and contract data, including pricing, margins, and terms a competitor would value.
Staff information such as performance notes or salary figures pasted in for a quick redraft.
How to see and control what ChatGPT remembers
You can manage all of this inside the ChatGPT settings, and it takes about ten minutes to set a sensible baseline on one account. Then walk your team through the same steps.
Open Settings, then Personalisation, then Memory, and open the memory summary to read what has been stored about you.
Delete any entry that holds client, financial, or staff information. Removing an item stops it shaping future answers.
Turn off 'Reference saved memories' and 'Reference chat history' if you want ChatGPT to stop building a profile at all. This suits accounts that touch sensitive data.
Use a Temporary Chat for one-off tasks involving client information, since temporary chats are kept out of memory and history.
Check the separate data control that decides whether your conversations train future models, and switch it off when your data is confidential.
That last point catches many teams out. The memory control and the model-training control are two different settings. Turning memory off does not automatically stop your conversations being used to improve the model, and clearing memory does not pull back data already processed. Treat them as separate decisions and set both.
A simple memory policy for your team
Settings on one laptop do not scale, but a one-page rule everyone follows does. The cost of getting this wrong is real. A single reportable breach investigation can absorb $45,000 in legal and remediation time before you count the damage to client trust, and serious or repeated breaches now carry penalties reaching into the millions. A short policy is cheap insurance against both.
Default to a Temporary Chat whenever a task involves a named client, patient, or candidate.
Never paste whole documents that contain personal or financial data. Redact the identifiers first.
Keep memory and training controls switched off on any account used for client work.
Review the memory summary monthly and clear anything that should not be sitting there.
Where Claude fits for governed deployments
For casual research and drafting, the controls above are enough. Once AI becomes part of how your business actually runs, the questions get sharper: where does the data sit, who can see it, and can you prove it. This is where we usually point Sydney clients toward Claude. Anthropic builds its Claude team and enterprise plans so that business inputs and outputs are not used to train models by default, which removes one of the harder conversations that consumer ChatGPT memory forces on you.
We help Australian businesses run AI that is useful and defensible at the same time, from memory and data settings through to a written policy your staff will actually follow. If you want a second set of eyes on how your team uses ChatGPT, Claude, or both, book a short brainstorm and we will map the gaps with you.
