ChatGPT Memory Not Working? Check Settings, Scope, and Sources
When ChatGPT misses something you expected it to remember, check the account, memory controls, conversation type, and source. Use one small test to find the gap before changing more settings.
Updated 6 min read
When ChatGPT memory is not working as expected, check the account, available memory settings, conversation type, and source of the missing detail. Then test one harmless fact in the intended scope before resetting settings or deleting useful records.
For work that cannot depend on uncertain recall, fenbs keeps a deliberate project recap and Where we left off outside one chat. That will not repair a ChatGPT setting, but it gives the next supported connected assistant an explicit starting point.
What should you check in order?
- Confirm the account and workspace. Check that you are testing where you originally supplied the information, especially if you use both personal and managed accounts.
- Open Settings, Personalization, then Memory. Note which controls your account actually exposes and whether the relevant ones are enabled.
- Identify the conversation type. An ordinary chat, temporary conversation, and project chat can have different context boundaries.
- Find the original fact. Inspect the source or saved record rather than assuming an earlier “I’ll remember” response established durable recall.
- Run a small test with one harmless fact, then compare the answer with the intended source. Change one condition at a time.
OpenAI explains the available controls in Memory in ChatGPT (opens in a new tab). Memory is selective, and its interface varies by account. A missing detail does not by itself prove that the whole feature has failed.
Write down the expected behavior precisely: “A new ordinary conversation should use my current preference for concise answers.” That is testable. “It should remember my project” leaves unclear whether you mean a preference, a file, an approval, or an unfinished task.
Was the information available to this conversation?
Locate where you supplied the detail and what kind of detail it was. You may have written a fact only in a separate project, left it in a file that the current chat does not have, or discussed it as a possibility rather than an approved instruction.
| Symptom | Check first | Useful evidence |
|---|---|---|
| The setting is missing | The current account, workspace, and available controls. | Which settings are visible and which workspace you are using. |
| A preference is absent | Whether the remembered record or direct instruction exists. | The current wording of the preference and where it is stored. |
| A project fact is absent elsewhere | The project’s context boundary. | The source conversation or document and its project. |
| An old fact keeps returning | Whether current and outdated sources conflict. | The current approved answer and the older source it replaces. |
| A task resumes at the wrong step | Whether progress was actually recorded. | Completed work, remaining checks, and the next action. |
For example, suppose a content designer discussed two possible audiences, but only later chose one. A new response aimed at the wrong audience is not fixed by repeating “remember better.” Supply the approved choice, identify the earlier option as rejected, and keep that distinction in the project record.
When the problem is missing work context, a short fenbs project record can state the audience, the person who chose it, and the next drafting task. The assistant then has a concrete answer to read instead of an ambiguous conversation to reconstruct.
Could the conversation’s scope explain the result?
OpenAI’s Temporary Chat guide (opens in a new tab) distinguishes personalized and unpersonalized temporary conversations; neither creates or updates memory while temporary. Check the choice made at the start before testing whether a new detail persists.
For project work, check the project’s settings and source location. OpenAI documents the boundary in Projects in ChatGPT (opens in a new tab). A fact being available in one place does not establish that a conversation elsewhere should reference it.
Do not loosen a work boundary just to make a test pass. If the task needs an approved requirement, provide that requirement in the appropriate project. Keep private information out of a shared workspace unless it belongs there and you have permission to use it.
For adjacent problems, see general ChatGPT failures, project failures, Copilot recall, deep research failures, or scheduled task failures.
How can you run a useful recall test?
Use a harmless preference that is easy to recognize and relevant to the test, such as wanting a short summary before an explanation. Avoid client details, personal identifiers, or confidential material. The goal is to isolate the behavior, not to create a new privacy problem.
- Choose the intended account and an ordinary conversation with the relevant settings enabled.
- Provide the test preference clearly and inspect the saved record when that control is available.
- Start a fresh conversation in the same intended scope. Ask for an ordinary answer without repeating the preference.
- Compare the result with the saved wording. Note what happened, then remove the test preference if you do not actually want to keep it.
One successful answer does not prove permanent recall. One unsuccessful answer also needs interpretation: the request may not have made the preference relevant. Repeat only enough to identify a clear pattern, and keep the question comparable rather than changing the topic, account, and settings together.
If the missing context should come from an outside app, use the ChatGPT connection troubleshooting guide. If it comes from written development instructions, instruction troubleshooting examines a related but separate failure. Do not treat every missing source as a personal-memory problem.
What should you send to support if the test still fails?
Prepare a short report: the account type and workspace, the platform, visible memory controls, the intended source, the test steps, and the actual result. Include the time and time zone. Mask private material in screenshots and avoid sending passwords or authentication codes.
Use OpenAI Support’s contact instructions (opens in a new tab) for the current channel. If the setting is controlled by your organization, ask its administrator first. Describe the missing behavior instead of asserting a cause you have not established.
Keep a separate progress note while you investigate. “The draft exists, the audience is approved, and the citations still need checking” lets you resume useful work without repeatedly testing recall on a live assignment.
Keep progress readable while you resolve recall
Save the current task status and next action in fenbs Where we left off. Add approved constraints in Decisions and rules and supporting facts in AI context notes. A connected assistant can read that explicit record at the start, even when you choose a different supported app for the next session.
Create a free accountFree for one person · No credit card · Works with Claude, ChatGPT, Cursor and more
How to do this with fenbs
- Create a free account and add the active project that needs a reliable recap.
- Add the task you are trying to continue. Record what was completed, what was checked, and the specific next step.
- Write the approved constraints in Decisions and rules. Add only the background notes needed to understand the work.
- Connect a supported assistant using its setup guide. Ask it to read the project context, explain where it will start, and save progress before stopping.
fenbs is free for one person, no credit card. It does not repair ChatGPT memory or import your chat history. It is not the right fit if you only need a missing personal setting restored; its job is to keep a deliberate record of work across sessions.
How do you reduce repeat failures?
Before important work, ask the assistant to state the constraints and next step from the available record. Correct a gap before it becomes a draft, code change, or message someone relies on. After the task, record what changed so the next session starts from the current answer.
Keep the record brief and specific enough that you can check it yourself. The AI memory topic hub covers broader ways to maintain context; for this failure, the useful result is knowing which information was missing and how the next session will obtain it.