You will be able to check, correct and switch off what an assistant remembers between chats.
Priya asked her assistant for ideas for a team lunch, and every suggestion was vegetarian. She is not vegetarian. It took her a while to work out why. Months earlier, she had asked for help planning a dinner for a vegetarian friend, and the assistant had saved a note that she preferred vegetarian food. It had been quietly shaping food suggestions ever since.
That is memory at work. It can be genuinely useful, and it can also go wrong in ways you will only notice if you go and look at what has been saved, which is what this lesson shows you how to do.
Many assistants now have a memory feature. As you chat, the assistant saves short notes about you: your job, your family situation, your preferences, projects you are working on. In later chats, relevant notes are inserted into the conversation, so the assistant seems to remember you. Some products save notes automatically, some only when you ask, and some can also refer back to your past chats. Products differ, so check your assistant's help pages for how its version works.
The important point is what memory is not. The model itself is not learning about you. As AI fundamentals lesson 5.1, The context window is the model's only working memory, explained, the model's weights do not change while you chat. Memory is a system around the model that saves text and pastes it back in, much like the custom instructions in lesson 6.1, Custom instructions: what to say once. The difference is that you wrote your custom instructions, and the memory notes were mostly written by the assistant, based on what it decided was worth keeping.
Because the notes are written for you, they can be wrong. A one-off request can be saved as a general preference, as with Priya's vegetarian friend. A joke can be taken literally. A detail you mentioned about a colleague can end up recorded as if it were about you.
Even accurate memories go out of date. If you changed jobs, moved house or finished a project, the old notes may still be there. Lesson 6.2, Projects: a workspace for each recurring job, warned that an outdated file in a project gets used confidently, and memory is no different. The assistant does not know that "works in HR at a logistics company" stopped being true last month.
The quiet part is what makes this a problem. You usually do not see which memories were inserted into a chat. The answer just leans in a direction, and unless the effect is obvious, as with a string of vegetarian restaurants, you may never notice. A stale note about your job might tilt every piece of career advice you ask for.
Most assistants that have memory give you control over it. You can usually view the list of saved memories, delete individual ones or clear them all, and switch the feature off entirely. Look for it under settings, often in a section called memory, personalisation or something similar.
Make a habit of reviewing it. When you open the list, you are looking for three things: notes that are wrong, notes that are out of date, and notes that are accurate but that you would rather the assistant did not keep, such as details about your health, your finances or other people, and delete every one of them.
Then decide whether to keep memory on at all. Some people find it saves time; others prefer the predictability of an assistant that only knows what they put in custom instructions, projects and the chat itself, and either choice is reasonable. If you keep it on, a review every month or two catches most problems before they build up.
Many assistants also offer a temporary chat, sometimes called an incognito or private chat. These are usually kept out of your chat history and do not create memories, and in most products they do not draw on existing ones either. The details of what is kept, and for how long, vary between products and are set out in their privacy pages.
Use one for anything sensitive or one-off that you do not want carried forward: a health question, a question about a family member's money, a draft of a difficult personal message, or a request you know is unusual for you, like Priya's dinner for her vegetarian friend. A temporary chat keeps that out of your memory and away from every future answer.
A temporary chat is not a privacy guarantee. It controls what the assistant remembers about you, not what you are allowed to share. Your employer's rules and the product's data settings still apply, which lesson 8.2, Data settings and the rules at your workplace, covers.
You may be surprised by what is already saved. Find the memory or personalisation section in the settings of the assistant you use most, and have a pen ready, because you will want to note what you find before you start deleting.
Open your assistant's memory or personalisation settings, list what it has saved about you, and delete anything wrong or that you would rather it did not keep.
Junxiong-WFG Organisation is an authorised representative of AIA Financial Advisers Private Limited (Reg. No. 201715016G).