Field noteReport a correction

Why Does ChatGPT Remember Other Chats? Memory, Chat History, and Temporary Chat

SOURCED

Metadata

  • Type: Research Record
  • Working title: Why Does ChatGPT Remember Other Chats? Memory, Chat History, and Temporary Chat
  • Working slug: /en/research/why-chatgpt-remembers-other-chats/
  • Search intent: why does ChatGPT remember other chats / ChatGPT memory after deleting chat / does deleting a chat delete memory / ChatGPT Temporary Chat
  • Tags: Memory, Chat history, Privacy, Personalization, Research record
  • Environment: ChatGPT on web and mobile. Controls and UI may vary by plan, rollout, and memory settings.
  • Last checked: August 20, 2026

Short answer

If ChatGPT brings up something from another conversation, that is not evidence that it read a draft you never sent. When memory is enabled, ChatGPT may personalize a response using context from past chats, saved or summarized memories, uploaded files, and connected apps.

Deleting a chat and deleting a memory created from that chat are not the same action. OpenAI's current Memory FAQ says that fully removing a fact may require deleting every source in which it appears, including past or archived chats, files, the memory summary, and connected apps.

What ChatGPT may reference

OpenAI says memory can use context from chats, files, and connected apps. Where the feature is available, the book icon below a response can show sources used for personalization, such as custom instructions, past chats, files, and memories.

That display is not a complete audit log. OpenAI also says the memory summary may not show everything ChatGPT remembers, and the sources panel may not reveal every factor that shaped a response.

Project memory can now be changed in eligible projects

OpenAI's release notes dated August 14, 2026 say eligible unshared projects can switch between default memory and project-only memory after creation. With project-only memory, ChatGPT may use conversations from the same project, but it should not reference memories or chats outside that project, and project information should not feed memory used in outside chats.

Shared projects remain project-only and cannot be switched. OpenAI also notes that a change may take a few hours to apply. A single answer immediately after changing the setting is therefore weak evidence; check the visible memory mode, whether the project is shared, and whether the change has had time to propagate.

Separate the possible sources of this answer into: (1) this chat, (2) other chats in this project, (3) saved memory, (4) other chat history, (5) files or connected apps, and (6) general knowledge. If you cannot verify a source, label it “not verifiable” rather than guessing.

Limit: the model's self-report is not an audit log. Where available, compare it with the project's visible memory setting and the Sources shown beneath the response.

What this does not prove

  • A close match does not prove that ChatGPT read text deleted from an unsent message box.
  • Asking the model where it learned something is useful for triage, but its answer is not forensic evidence.
  • A missing item in the memory summary does not prove that earlier context had no effect.
  • The model may simply have guessed from the current conversation or produced a common answer that happened to fit.

Why it can feel paranormal

People often treat a new chat as a first meeting. The product may treat memory and account-level history as continuing context. That mismatch can make ordinary personalization feel like the system knows something it was never told.

A public Hermes Agent issue documents a related mechanism: irrelevant stored memories were injected into context and caused an unrelated request to be interpreted as a continuation of an older project. The report involved Hermes Agent v0.13.0 with MiniMax-M2.7, not ChatGPT, so it is evidence of a general failure mode rather than proof of ChatGPT's implementation.

Practical context-check prompt

Did this answer use information from outside the current chat? If it may have, separate the possibilities into: (1) saved or summarized memory, (2) past chats, (3) uploaded files, (4) connected apps, and (5) inference from the current conversation. Write “cannot verify” wherever you cannot determine the source.

Use case: identifying plausible sources of unexpected personalization.

Limitation: the model's self-report is not an audit log. Check the memory sources shown below the answer, when available, and review the memory summary in settings. For a safer comparison, use harmless test facts and compare a normal chat with Temporary Chat.

Facts, reports, and editorial inference

Verified facts

  • OpenAI's current FAQ says memory may use context from chats, files, and connected apps.
  • The memory summary and the sources display are not necessarily exhaustive.
  • Temporary Chat does not use existing memories or create new ones.
  • Fully removing a fact may require deleting it from every source where it appears.

User reports

  • A Japanese Zenn author reported that context from earlier analysis and a personality test appeared in unrelated chats. The post does not provide a controlled third-party reproduction.
  • A Hermes Agent issue reports unrelated stored memories steering a new request toward an old project. It concerns another product and model.

Editorial inference

  • The mismatch between “new thread” and “new relationship” is a major source of the uncanny effect.
  • Memory retrieval, ordinary inference, coincidence, and selective reporting can all produce the appearance of secret knowledge.

Primary sources and scope