CYBER WELFARE

Protect your Digital Privacy

Signs your AI assistant remembers too much: how to notice and what they mean

An artificial intelligence assistant’s memory works quietly. It picks up a detail in one conversation, reuses it in another, and usually does not announce it. When it saves you repeating who you are and what you do for a living, that is exactly what you want. When it brings into an answer something you do not remember saying, it is worth noticing: it is the simplest way to tell whether the settings are deciding for you.

This post gathers the signs your AI assistant remembers too much: some you find in the settings, others you notice in the answers. None of these signs, on its own, means something has gone wrong. Together, they tell you whether it is time for a tidy-up.

It is the diagnostic side of the recommendation to check your AI assistant’s data settings. The ways other people may target chats, memories and connected apps are explained in the post on prompt injection attacks.

What these signs are

In cyber security people talk about indicators: traces suggesting that a situation deserves attention. With an AI assistant, indicators come in two kinds:

  • technical, meaning information the service shows in its settings: the memory, the history, connected apps, shared links, devices signed in to the account;
  • observable, meaning behaviour you notice while using the assistant: an answer that knows too much, a wrong detail that keeps coming back, an action you did not ask for.

An indicator is not proof. It is an invitation to look more closely.

A rule already used elsewhere applies here too: a weak signal on its own is not proof; a strong signal, or two together, deserve action. An answer that remembers your job is a weak signal. A line in the memory you do not recognise, together with a connected app you do not remember, is already a good reason to stop and check.

Why they matter more if you use the assistant for many things

Someone who uses an assistant for the odd question leaves few traces. Someone who uses it every day, for work, study and private life, with memory on and a few connected apps, builds up a detailed profile within months: projects, people, habits, places. The more areas of life share the same account, the more likely it is that a detail born in one context turns up in another.

Services also differ. On some, the memory updates itself, collecting whatever seems useful; on others, it saves only what you ask it to remember. Knowing how yours works is the first check.

How you use your devices matters too. An assistant left open on the family tablet, the office computer or a phone you lend to a friend shows your history and answers with your memory to whoever uses it. Many signs start right there: someone else sees an answer that talks about you, and tells you.

Technical indicators

IndicatorWhere you see itWhat it may meanHow much it weighsWhat to do
Memory with many lines you do not remember givingIn the settings, items such as memory or personalisationThe assistant saves details from conversations by itselfMediumReread it and delete what you do not want kept
Memory lines about other peopleList of memoriesDetails about family, colleagues or friends have become “yours”Medium to highDelete them
Lines you do not recognise at allList of memoriesA summarising error, or an instruction that came from outside contentHighDelete them and check connected apps and sign-ins
Long history, no auto-deleteIn the settings, items such as history or activityChats stay until you delete themLow to mediumChoose a period
Connected apps you do not rememberIn the settings, items such as connected apps or connectorsA trial connection still activeHighDisconnect them
Shared links still liveList of links, where the service shows oneConversations readable by anyone with the linkHighRevoke them
Unknown devices or sessionsAccount settingsSomeone may be using your accountHighSign out of the sessions and change the password

Signs you can spot yourself

SignWhen you notice itWhat it may meanHow much it weighsWhat to do
An answer mentions a detail you did not write in that chatIn a new conversationMemory, or reference to past chats, is onLow to mediumCheck the memory
A wrong detail comes back in several answersOver timeA wrong memory has been savedMediumCorrect or delete it
The assistant mixes work and private lifeIn a work answerOne account gathers different areasMediumSeparate the uses or review the memory
It offers to continue a sensitive topic you had closedOn openingThe service reuses history for suggestionsMediumDelete the chat and the related memories
An action on email or calendar you did not ask forAfter it read a page or documentOutside content may have given the assistant instructionsHighDisconnect the app and check what was done
Someone tells you they have read one of your conversationsFrom a contactA shared link has been passed aroundHighRevoke the link
Whoever uses your device sees answers personalised about youOn a shared deviceAccount and memory are the same for everyoneMediumUse separate accounts or the temporary chat

Signs on other devices and in other programs

The assistant does not live only in its own app. It is often built into the browser, the phone keyboard, a writing program or an email service. Some signs are typical of these situations:

  • an assistant built into another program answers with details you gave elsewhere: the same account can link several services;
  • an AI feature switched itself on after an update: its data settings need checking too;
  • on the family computer the assistant is already open with your account: anyone who uses it sees your history;
  • a browser extension asks to read every page you visit to “improve AI”: that is a broad permission, worth weighing up calmly.

In all these cases the check is the same: open the data settings and ask, for each function, whether it matches a choice you have made.

A concrete example

Mark is 26 and uses an assistant for everything: his CV, recipes, the gym, questions for a Spanish exam. One evening he asks for an idea for a dinner with friends and the answer begins: “Since you are looking for a job in Manchester and have just split up with Julia, why not…”. It is all true. But Mark never wrote those two things in the same conversation, and had not realised the assistant had put them together.

He opens the memory and finds thirty-two lines. Some are useful, like his level of Spanish. Others are not: his ex’s name, the area he lives in, the fact that he is flat-hunting, and one wrong line saying he has “two children”. He also looks at the connected apps: his calendar is there, connected for a trial months earlier.

Mark deletes the lines about his ex, about the flat and the wrong one, and keeps the ones about work. He disconnects the calendar. Then he decides to use the temporary chat for personal matters.

Nothing serious happened. But the sign was there: an answer that knew more than Mark had decided to say.

The following week Mark checks again. There are two new lines in the memory, both about work and both correct: he keeps them. He notices that the service lets him see when a memory is added, and decides to have a look whenever he talks about something personal. Since then, the check has taken him a minute.

What to check straight away

If one of the signs in the tables sounds familiar, these are the checks to do first, in this order:

  1. Are there connected apps or shared links you do not recognise? They come first, because they open access to others.
  2. Is the account signed in on devices you do not know? If so, sign out of those sessions and change the password.
  3. What is in the memory? Reread it and delete wrong details, details about other people and anything you do not want kept.
  4. Does the history have a set period? If not, choose one.
  5. Is training your choice? Check the setting, knowing that it covers future conversations.
  6. Do you use the same account for work and private life? Consider separating them.

The full sequence for putting things in order, from prevention to recovery, is in the guide on how to delete AI chat history and memory.

If you find a suspicious sign

Stay calm: in most cases it is simply a memory that is more active than you thought. If, however, the sign concerns an action on email or calendar you did not ask for, a memory you do not recognise at all or a sign-in from an unknown device, it is worth:

  • disconnecting the connected apps and checking sent emails and changed appointments;
  • changing the account password, using one you do not use anywhere else, and turning on two-step verification;
  • deleting the memories you do not recognise;
  • if the account is a work one, telling your organisation’s IT team.

Over the following days keep an eye on two things: sign-in alerts from places or devices you do not recognise, and unusual messages in your name reported by your contacts. If nothing turns up, it is very likely there were no consequences. The typical signs of a connected service behaving oddly are also collected in the post on the signs of suspicious connected apps.

What is not an indicator

Some situations look suspicious but are usually normal:

  • the assistant remembers your name or your job: if you told it and memory is on, that is what it is for;
  • an answer picks up the conversation from an hour ago: on many services the assistant can refer to recent chats, if the option is on;
  • the temporary chat does not appear in the history: that is exactly what it should do, even though the service keeps it for a short period;
  • after turning training off, your chats stay in the history: on many services the two settings are separate;
  • a suggestion based on what you have just written: within the same conversation the assistant uses everything above, and it is normal for it to take that into account;
  • a general answer about your town: the assistant can infer the area from your connection, without it being a memory.

The way to tell the difference is always the same: a memory is normal when you gave it and know it is there; it is a sign when you do not recognise it. If in doubt, the cautious choice costs nothing: open the memory and look.

How often to check

MomentWhat to check
When an answer knows more than you expectedThe memory and recent history
When you connect a new appThe permissions granted and the apps already connected
Once a yearMemory, history, training, connected apps and shared links
When the service announces a new featureThe data settings, which may have changed

Two important caveats

First caveat: no sign is conclusive. An assistant can remember a lot without anyone having done anything wrong, and it can behave normally even when a connected app has too many permissions. The most reliable check remains the simplest: open the settings regularly.

Second caveat: menus change. Feature names and settings paths vary between services and versions, and they often change with updates. If you cannot find an item, search the service’s official help pages for words such as “memory”, “personalisation”, “activity” or “connected apps”. And if a new feature appears after an update, treat it like a new assistant: before using it, see what it keeps.

How this connects to the Cyber Welfare Framework

PillarWhat this content contributes
AwarenessRecognising that an answer that “knows too much” is a sign, even when it seems harmless
SkillsBeing able to read the memory, the list of connected apps, shared links and the account’s devices
Secure BehaviourStopping at unknown memories or actions you did not ask for, and checking the settings regularly

Reference level: FL2 — Beginner. This is the level at which you start noticing what the assistant does by itself and asking why. To see where you stand on your other digital habits, take the digital resilience self-assessment.

Conclusion

The most common sign is also the simplest: an answer that knows more about you than you remember saying. In most cases it is the memory doing its job. Sometimes it is a wrong memory, a forgotten connected app or outside content that gave instructions. In every case the response is the same: open the settings, look and decide.

Noticing is the first step. The second is making sure the assistant remembers only what you have decided.

Related resources

Short reads from the Resources section, for anyone who wants to focus on a single aspect:

Related content

Start with the first step: the Cyber Welfare Programme guides you free of charge, one recommendation at a time.