Konshus
AI Memory News

Why AI assistants forget you between chats

A saved conversation is not the same thing as a remembered one. Across major assistants, recall depends on settings, selection, retrieval and whether the model uses the information it receives.

2026-11-11 · The Konshus Team · Sources checked before publishing

Share

Want more news like this?

  • One short email a week with the biggest changes in AI memory — ChatGPT, Claude, Gemini, new companies and research.
  • Every story is checked against its sources before it goes out.
  • No spam: just the weekly news, no sales emails, and we never sell or share your address.

Free. One email a week, nothing else — no Konshus marketing emails, and we never sell or share your address. Unsubscribe anytime in one click.

You told an assistant how you like to work. The next day, in a new chat, it asks again. Did it forget, or was it never given the detail in the first place?

The answer depends on a distinction that product interfaces can make hard to see. A conversation can be stored in your account without being used in the next conversation. A memory feature may select information from earlier chats; a separate step must then make relevant information available to the model answering you. OpenAI says ChatGPT does not search chat history for every response. It looks for relevant context when that is likely to improve the answer. OpenAI’s memory FAQ makes the practical consequence clear: seeing an old chat in a sidebar does not mean its contents were supplied to a new one.

What changed: more ways to remember, not a promise to remember everything

Cross-chat recall is no longer one feature with one obvious switch. ChatGPT describes saved memories and, where available, reference to past chats. Claude describes memory across chats as well as a way to search earlier conversations when prompted. Gemini distinguishes information a user saves from its option to reference past chats. Microsoft describes Copilot personalization and memory controls separately from the ability to view past conversations. OpenAI, Anthropic, Google and Microsoft document those distinctions.

That is useful progress. These products offer ways to carry context forward rather than making people repeat every preference or project detail. But their documentation describes different controls and boundaries, not one shared standard for what an assistant must recall. The sensible question is less “Does this assistant have memory?” than “Which information can this particular chat use, and how does it get there?”

First, check whether cross-chat use is available

A new chat may start without personal context because the relevant feature is off or unavailable for the account. OpenAI says ChatGPT memory settings and availability can vary by plan, region, platform and workspace; a missing setting may mean it is not yet available to that user. OpenAI’s memory FAQ is a better guide to that uncertainty than assuming every account behaves alike.

Google sets explicit conditions for Gemini’s past-chat personalization, including an eligible personal Google Account, an age of at least 18 and Keep Activity being on. It says the feature is not available for work, school or supervised accounts. Google’s eligibility guidance illustrates why two people using an assistant with the same name may have different experiences. Microsoft says turning off Copilot personalization and memory stops future personalization while leaving past conversations viewable. Its privacy controls likewise separate access to a record from permission to reuse it.

The first troubleshooting step is therefore mundane: inspect the settings for the account and workspace you are actually using. A transcript on screen answers “Was this saved?” It does not, by itself, answer “Can this new chat use it?”

Then, ask what was selected

Even with memory enabled, a system need not preserve a perfect account of everything you said. OpenAI says ChatGPT does not retain every detail from every conversation and that what it remembers can change as context changes. Its FAQ describes a selective feature, not a permanent transcript copied into every response.

The selection can take different forms. Anthropic says Claude saves memory as individual topics while you chat rather than writing a summary of each conversation after it ends. Claude’s memory guidance describes that approach. Google, meanwhile, distinguishes facts or preferences saved for Gemini from its separate past-chat reference feature. Gemini’s help page explains both options.

For a stable preference you expect to matter later, an explicit request to save it is more dependable as an intention than leaving it buried in an exchange and assuming it will become memory. That is not a guarantee of future recall. It is a way to make the request visible and, where the product offers controls, easier to check.

A memory can exist without being used

Selection is only part of the problem. Information may be available somewhere in a product without being retrieved for the question in front of it. OpenAI says ChatGPT looks for relevant past context when it expects that context to help; it does not search history for every answer. Its documentation leaves room for a detail to matter to you but not be picked up for a particular reply.

Some products provide a more directed route. Anthropic says users can ask Claude to search previous conversations for relevant details. Google says users can ask Gemini to refer to past chats and that Gemini may identify “Previous chats” among a response’s sources when it uses them. See Claude’s help page and Gemini’s help page. If the earlier detail matters, asking about the particular project or conversation is more informative than asking vaguely, “Do you remember me?” A search request cannot, however, overcome an unavailable feature or an inaccessible chat.

There is also a limit after retrieval: the model must use what it receives. In the Lost in the Middle study, tested language models often performed better when relevant information appeared near the beginning or end of a long input than when it appeared in the middle; performance also fell as inputs grew longer. The researchers’ paper examines particular models and tasks, not every current assistant. It nonetheless cautions against treating a long context window as proof of dependable recall.

Boundaries and privacy choices can look like forgetfulness

Before blaming retrieval, check where the earlier conversation happened. OpenAI says ChatGPT temporary chats do not create or update memories. Its memory FAQ describes a deliberate exception to persistence. Anthropic says each Claude project has its own memory: a project can use its memory alongside account memory but not memories from other projects. Claude’s project documentation makes that boundary explicit. A detail supplied in another project, account or setting should not be assumed available here.

Forgetting can also be the intended result of a privacy choice. OpenAI says turning off ChatGPT’s Reference chat history schedules information remembered from past chats for deletion from its systems within 30 days, while the original chats remain unless separately deleted. It also says saved memories are stored separately, so deleting an original chat alone does not necessarily delete its saved memory. Those distinctions appear in its FAQ. They cut both ways: a chat may remain visible without personalizing future answers, and a saved memory may need separate attention after its source chat is removed.

What this means for a fresh chat

If an assistant misses something important, work through the layers rather than treating “memory” as a single yes-or-no ability. Check whether cross-chat personalization is available and enabled. Check whether the information was intentionally saved, merely mentioned, or confined to a temporary chat or separate project. Where the product supports it, ask for a specific search of earlier conversations. For a consequential detail, provide it again and verify that the answer uses it correctly.

What remains unsettled is not whether companies can build useful recall features—the documented controls show that they can. It is how reliably any given assistant will select the right detail, retrieve it for the right question and apply it without bringing forward something a person meant to leave behind. Until those steps are clearer to users, “It forgot me” will describe several different failures, and sometimes a privacy boundary working as intended.

Our take (opinion)

The industry has made real progress by separating saved information, past-chat reference and conversation search. Those are useful choices, and it is fair to credit products that give people controls rather than making every earlier remark automatically available.

But “memory” remains too broad a label for a user trying to diagnose a mistake. A visible transcript can create an expectation of recall that the product never promised. Conversely, deleting a transcript may not remove a separately saved memory. Our view is that assistants should make these boundaries legible at the moment they matter: what was saved, what was brought into this answer and which control changes future use. That would help people distinguish a retrieval miss from a disabled feature or an intentional privacy limit. Until then, the safest habit is to state important context again and check the result—not because cross-chat memory is useless, but because selective recall is not the same as a dependable record.

Questions

If I can see an old chat, why can’t my assistant use it?

Keeping a transcript and using it to answer a new chat are separate functions. Cross-chat use depends on the product’s features, settings, eligibility and whether relevant information is selected or retrieved.

Will asking an assistant to remember something fix the problem?

An explicit request can help where saved-memory features are available, particularly for a stable preference. It does not guarantee that the detail will be retrieved or used correctly in every later answer.

Does deleting a chat delete what the assistant remembers?

Not necessarily. OpenAI says ChatGPT saved memories are stored separately from chat history, so deleting the original chat alone may leave a saved memory. Check the product’s separate memory and history controls.

Does a bigger context window mean an assistant will remember more reliably?

Not by itself. Research on tested language models found that relevant information in long inputs could be harder to use when it appeared in the middle. That finding does not establish how every current assistant performs.

Sources

  1. Memory in ChatGPT | OpenAI Help Center
  2. Use Claude's chat search and memory to build on previous context
  3. How can I create and manage projects? | Claude Help Center
  4. Save info & reference past chats in Gemini Apps
  5. Get personalization with memory of your past Gemini chats
  6. Microsoft Copilot privacy controls
  7. Lost in the Middle: How Language Models Use Long Contexts

The AI memory news, once a week.

If you're interested in AI memory, want your own AI to remember you better, or just want to stay up to date, this is the newsletter for you. Every week we comb through the news for the big shifts in AI memory — new companies, new features from ChatGPT, Claude and Gemini, new research and new ideas — and sum up what matters in one short email.

For people who use AI every day, builders and founders, and anyone curious where AI memory is going.

Free. One email a week, nothing else — no Konshus marketing emails, and we never sell or share your address. Unsubscribe anytime in one click.

About Konshus: a private memory Vault that works across ChatGPT, Claude, Gemini and more. Learn more.