Memory, and the notes it took about you

Something is writing things down

Several products now keep persistent notes about you across conversations. ChatGPT calls it memory, Gemini has saved information and personal context, Claude has project knowledge and, in some configurations, conversation history it can search. The mechanism is the same in each case: facts get saved somewhere, and relevant ones get injected into the prompt before the model answers.

This is useful. It is also the layer people understand least, and it produces the strangest bugs, because the injected text is invisible in the conversation you are looking at.

The good part

You stop repeating yourself. It knows you write in British English, that you are preparing for a viva in molecular biology, that your daughter is allergic to peanuts, that you work in Kenyan shillings. Answers arrive already fitted to you.

For most people the value is real and it accumulates over months.

Four ways it goes wrong

A one-off becomes a permanent fact. You asked once, for a friend, about a diet plan. Three months later your recipe suggestions are all low-carb and you have no idea why. The saved note said you were interested in low-carb eating. It was true for four minutes.

A stale role sticks. You changed jobs. The memory has you as a teacher, and every explanation still arrives pitched at a classroom.

It contaminates a fresh start. The whole point of starting a new conversation, when an old one has gone wrong, is a clean context. Memory can quietly carry over the very assumption you were trying to escape. If you restart and get the same wrong answer, check memory before concluding the model is stubborn.

It goes places you did not intend. Anything saved can appear in a later conversation, including one you are screen-sharing, one that produces a document you send to a client, or one somebody else is watching. A saved note about a health condition or a difficult colleague is now in a system that will use it when it seems relevant.

The maintenance, which takes five minutes

Every product with memory has a page where you can see and delete what it stored. Find yours and read it once, properly.

• ChatGPT: Settings → Personalization → Manage memories.
• Gemini: Settings → Personal context / Saved info.
• Claude: Settings → Capabilities, where memory can be turned on or off; project knowledge is visible in each project.

Names and menus move; the thing to look for is a list of remembered items, not a toggle.

Most people are surprised. Not usually by anything sinister: by how much is trivial, out of date, or the residue of a question they asked on behalf of somebody else.

Then adopt three habits:

1. Delete stale items when your role, location or project changes.
2. Use a temporary or incognito chat for anything you do not want retained: asking on behalf of a friend, exploring a sensitive question, testing something you do not want influencing future answers. Every major product has one.
3. Say it explicitly when you want something kept: “remember that our financial year ends 31 December”. Do not rely on it inferring which facts are durable.

Memory versus a standing brief

They are not substitutes, and the difference is worth stating plainly.

A standing brief is a document you wrote. You can read it, edit it, share it with a colleague, and reason about it. Memory is a set of notes taken by a system about you, whose contents you did not choose and whose selection into any given prompt you cannot see.

For anything that matters (professional context, hard constraints, house vocabulary) write the brief. Let memory handle the convenience layer. Somebody who relies on memory for the important context ends up unable to answer the question “why did it say that?”, which is a bad place to be when the answer went to a client.

The one-line test

If an answer surprises you, ask: what is in this prompt that I did not put there? System prompt, retrieved passages, tool results, and memory. It is the one people forget, and it is the one that changes between you and the colleague getting different results from the same words.