Chatbot memory: what they really remember

A caterer asks his assistant to draft a reply to a client. The text mentions "a wedding of forty people in the Dordogne." He never wrote that in this conversation. Three weeks earlier, he had asked for a quote for that wedding, in another discussion. Either the tool remembers it, or it just made it up, and he needs to know which.

"Memory" covers three different things. Confusing them is the source of most surprises.

Three memories, three durations

The current conversation. This is the simplest. The tool rereads the whole exchange with each answer. Nothing is "learned": the text is simply in front of it. This memory has a size limit, the context window, and as the exchange grows longer, the beginning may be poorly taken into account. That's why a long discussion "drifts."

The workspace. Many tools let you group conversations into a project, with shared files and instructions. The tool rereads these elements at the start of each new discussion in the project. You choose what's in there, so you know what it knows.

Memory between conversations. Some assistants record information about you (your job, your preferences, your style) or draw on your past discussions to respond. You may then find, in a new conversation, a detail mentioned elsewhere. It's useful, sometimes unsettling, and it's what the caterer experienced.

These features differ by tool, plan, and country, and they change often: open your tool's settings to see what's turned on.

How to know what the tool has kept

Don't guess, ask, then check in the settings.

List everything you have retained about me from one conversation to the next, as a list, without adding or rephrasing anything. If you don't know what is stored, say so.

A chatbot's answer about its own memory can be incomplete or inaccurate. The official list is in the application's settings, usually with the option to delete an entry, erase everything, or turn the feature off. For a simple test, open a fresh conversation and ask a question that only makes sense if the tool knows you.

What's useful, what gets in the way

Useful: your job, your usual tone, the way you present a result. These are stable elements, of the same kind as what goes into system instructions.

What gets in the way:

• False or outdated memories. The tool keeps "I'm looking for premises" when you've already found some. It will keep bringing it up.
• Mixed contexts. A private detail shows up in a professional text, as with the caterer, or two clients contaminate each other.
• Skewed tests. If the tool knows you, it adapts its answers, and you're no longer measuring what your prompt is worth on its own.

Two habits take care of most of it. For work that has to be clean (a test, a document for a client), open a fresh conversation or use the mode that keeps nothing, when the tool offers one. And whatever matters for the task, write it in the prompt rather than counting on recall: the context to provide remains your responsibility.

What you shouldn't let it remember

Everything that's stored is data about you, kept by a third-party service under its own rules. Avoid letting health, identity, or financial information, or data about other people, get recorded. Tell the tool plainly if needed ("don't save this") without assuming that's enough, and then check the settings. Retention rules are described in each tool's terms, and the article on privacy gives you the reference points.

On a work account, your organization may have disabled or restricted memory.

To do this week

Open the settings of your usual assistant. See whether it keeps information between conversations, read the list, and delete whatever is wrong or whatever you don't want recorded. Then write three lines about yourself and your work in a separate file: you'll be able to paste them at the top of any conversation, whatever the tool.