AI Mixing Memory Between Your Projects? How to Stop It

You ask for feedback on a proposal for one client, and the tone comes back wrong. Not badly written, just casual in a way that does not suit this client at all. You reread your prompt. Nothing in it asked for casual.

What happened is that an assistant with memory on picked up a preference from a different piece of work, weeks ago, for a completely different client, and applied it here. Nothing on screen told you. You only notice when the output feels off, and sometimes not even then, which is the part that costs you.

This is the least discussed failure of AI memory, because it does not look like forgetting. It looks like the assistant being confidently unhelpful. There are three causes, and they have different fixes.

What it looks like when it happens

The signs are quiet. A tone that belongs to another account. An assumption from a project you closed in July. A recommendation that makes sense only if you are still doing the thing you stopped doing. Writers notice it first, because voice is the most obvious thing to contaminate: a light comedic register from one story turning up in a serious one.

The common thread is that the output is competent and wrong at the same time, and nothing flags it. If you have ever fixed something an assistant produced without being able to explain why it was off, this may be why.

Cause 1: your memory has no walls

By default, memory in most assistants is global. One pool, drawn from everything you have ever said, available to every new chat. It has no concept of client A and client B, or of work and personal, because you never told it there was a difference, and there is no place in the interface where you would have.

This is why the problem shows up most for people juggling several accounts or several projects. The more varied your work, the more likely that something learned in one corner gets applied in another.

Cause 2: what it kept about you is wrong

Memory is a summary rather than a transcript, and summaries go out of date. OpenAI has said plainly that its earlier memory entries drifted over time and could contradict each other, the published example being marathon training stored next to a sprained ankle.

So the entry steering your output may not even be a mismatch of context. It may simply be stale: a preference you had in March, a role you no longer hold, a constraint that was lifted. It will keep being applied until you remove it, and it does not announce itself. What memory keeps, and why a summary behaves this way, is covered in what AI actually remembers.

Cause 3: the container is not doing what you think

Projects, folders and workspaces look like walls. Some of them are, and some are only a way to group chats.

The difference is a setting rather than a feature, and the default is usually the permissive one. A project that has not been switched to isolated memory can still draw on everything else you have told the assistant. A project that has been switched will not. Both look identical in the sidebar.

Different assistants land differently here, and it is worth knowing where yours sits. What each assistant keeps goes through them one by one.

The setting most people never turn on

For ChatGPT, open the project’s own settings and look for the memory option. With project only memory on, chats inside that project ignore your saved memories and cannot reference conversations outside it, and conversations outside cannot reach in. It is available on the paid personal plans, and on business plans when memory is enabled for both the workspace and the account. Shared projects already work this way and cannot be switched back.

For Claude, each project keeps its own memory space by default, separate from other projects and from chats outside them, and you can read and edit entries individually.

For anything one off and sensitive, temporary or incognito mode neither uses existing memory nor creates new memory. It is the cleanest way to ask a question you do not want shaping future answers.

Three habits go with those settings. Open the memory page monthly and delete what is stale, since removal is the only way an entry stops being used. Start a genuinely separate piece of work in its own container rather than in whatever chat is open. And when an answer feels subtly wrong, check which sources it used before you argue with it.

When the fix is not a setting

Everything above works inside one assistant. It does nothing about the other half of the problem, which is that your work is probably spread across several.

Each one holds its own picture of you, built from whatever you happened to say there, invisible to the others and to you. Isolating projects in one app does not make the others consistent. It just gives you a tidy corner of a scattered room.

If that is your situation, the useful move is not another setting. It is to keep the material that matters somewhere you can see all of it, organised by project rather than by whichever app you were in, so what goes into a given conversation is something you chose rather than something that leaked. Which form solves which problem walks through the options, and for the occasional switch, moving context by hand still does the job.

One closing thought about why this particular failure is worth fixing first. Forgetting is annoying and obvious. Remembering the wrong thing is quiet, and you may act on it before you notice. More memory is not automatically better memory, and the difference is whether you can see and choose what it uses.