ChatGPT Memory can appear broken even when it is working normally. The important question is whether ChatGPT failed to remember something, failed to use an existing memory, or simply did not show that detail in the Memory Summary.
OpenAI changed Memory substantially in 2026. The newer system automatically builds and updates a broader understanding from useful context in chats and, where available, other sources. The Memory Summary is only a high-level view of that context, so a detail missing from the summary does not automatically mean ChatGPT forgot it.
Use a harmless test before changing anything:
- Open a normal, non-Temporary Chat.
- Tell ChatGPT a simple preference you are comfortable having remembered.
- Start a completely new normal chat.
- Ask about that preference.
Testing across two chats matters. Remembering something earlier in the same conversation can simply be normal conversation context rather than cross-chat Memory.
What Your Test Result Means
| Result | What to investigate |
|---|---|
| ChatGPT remembers it in a new chat | Memory is functioning |
| It remembers older context but not one detail | That detail may not have been retained or considered relevant |
| Memory Summary looks incomplete | The summary may not show everything Memory can use |
| Nothing carries into a new normal chat | Memory setting, account, workspace, or availability |
| Memory works normally but not in Temporary Chat | Expected behavior |
| Memory changes inside one Project | Project memory behavior |
| Memory disappeared after switching workspaces | Workspace settings or administrator controls |
This distinction prevents unnecessary steps such as deleting chats or clearing useful memory.
Check Whether Memory Is Enabled
Open:
Settings > Personalization > Memory
Make sure Memory is enabled.
OpenAI’s current Memory system can use useful context from your chats, files, and connected apps where supported. The exact sources and controls available can vary by account and workspace.
If you recently enabled Memory, do not expect a detailed Memory Summary immediately. OpenAI says a new account, a recently enabled account, or an account without enough chat history may initially show little or nothing.
Continue using normal chats and check again later.
An Empty Memory Summary Does Not Prove Memory Is Broken
This is one of the most important changes to understand.
OpenAI says the Memory Summary does not necessarily contain everything ChatGPT can remember or use. It is a high-level synthesis rather than a complete list of individual facts.
To see it, go to:
Settings > Personalization > Memory Summary > Manage
The summary can update automatically as ChatGPT learns which context remains useful.
If it looks stale, OpenAI also provides a Refresh option from the three-dot menu in Memory Summary management.
A refreshed summary still may not display every detail.
When you want to know whether ChatGPT remembers something specific, ask about that detail directly rather than using the summary as the only test.
Temporary Chat Is Supposed to Ignore Memory
Memory should not behave normally inside a Temporary Chat.
OpenAI says Temporary Chats do not use existing memories and do not create new ones.
That makes Temporary Chat useful when you want a conversation that does not affect personalization, but it also makes it a poor place to test whether Memory works.
Start a new ordinary chat and repeat the test there.
If Memory works in normal chats but not Temporary Chat, there is nothing to fix.
ChatGPT Remembered Something Before but Not Now
Memory is not designed as a verbatim archive of everything you have ever said.
The current system continually updates its understanding of what is useful. Some context can become less relevant, be replaced by newer information, or simply not be used in a particular response.
For example, ChatGPT may know a preference but not apply it when it has little relevance to the current question.
Ask directly:
What do you remember about my preference for [topic]?
If ChatGPT can describe the preference, the issue is not that Memory is completely unavailable. It may simply not have been relevant to the previous response.
For instructions that must be followed consistently, Custom Instructions may be a better place than relying solely on inferred memory.
Check Memory Sources on a Personalized Response
Current ChatGPT versions can show which personalization sources contributed to a response.
When available, tap the Sources or book icon below a personalized response.
OpenAI says sources can include items such as:
- past chats;
- memories;
- custom instructions;
- files where supported;
- connected-app context where supported.
This can answer a better question than “Is Memory on?”
It can show whether a response actually used remembered context.
Not every factor is necessarily displayed, so absence from the source panel should not be treated as a perfect audit of everything the model considered.
Memory Works in One Account but Not Another
Check which ChatGPT account you are using.
This matters when you sign in with different:
- email accounts;
- Google accounts;
- Apple accounts;
- personal and organizational workspaces.
A new or different account will not automatically have the same personalization context as the account you normally use.
Also check the active workspace.
Enterprise and other managed workspaces can have administrator-controlled Memory availability. OpenAI advises contacting the workspace administrator when a Memory setting is unavailable in an Enterprise workspace.
If Memory works in your personal workspace but not your managed workspace, changing browser permissions will not solve that difference.
Memory Behaves Differently Inside a Project
Do not use a Project as your only Memory test.
Projects can have their own context and memory behavior. A Project may emphasize its chats, files, and instructions differently from an ordinary conversation.
When the problem appears only inside one Project, test the same harmless remembered detail in a normal new chat.
If normal cross-chat Memory works, the account’s general Memory feature is not completely broken.
For Project creation or update failures themselves, see ChatGPT Error Creating or Updating Project: How to Fix It.
ChatGPT Does Not Remember Exact Text
Memory is intended for useful context and personalization, not reliable verbatim storage.
Do not expect it to preserve an entire:
- article;
- prompt library;
- legal document;
- spreadsheet;
- long template;
- codebase;
- password or secret;
- set of exact instructions word for word.
Keep important source material in an appropriate file or document and provide it when exact wording matters.
This distinction is particularly important for long-running work. Memory can help ChatGPT understand the project and your preferences, but the actual source document should remain the authority.
If you need ChatGPT to work directly from a document and attachments are failing, use ChatGPT File Upload Not Working? Find the Failure Point.
A New Memory Is Not Appearing Immediately
The newer Memory system is more automatic than the older saved-memory model.
A new detail does not necessarily become a visible standalone item in Memory Summary immediately.
OpenAI says the summary updates over time as ChatGPT synthesizes useful context.
For something important, state it clearly in a normal conversation. Then test it in a new normal chat.
Do not repeatedly add slightly different versions of the same preference simply because you cannot see an individual memory card. That can create unnecessary or conflicting context.
ChatGPT Remembers Outdated Information
This is a different problem from Memory not working.
Open Memory Summary > Manage and review the information shown there.
You can request a correction to the summary. Current ChatGPT also lets you highlight text in the Memory Summary to make a specific correction where that interface is available.
You can also correct ChatGPT naturally in conversation when an old preference or circumstance has changed.
For example:
I no longer use that workflow. Use my newer preference instead.
The improved Memory system is designed to keep context more current and reduce stale or contradictory information, but corrections are still useful when something has changed.
You Deleted a Chat but ChatGPT Still Knows the Information
Deleting one conversation is not necessarily enough to erase everything ChatGPT may know about a detail.
OpenAI’s current guidance says full removal can require deleting every source where that information appears, which can include:
- past chats;
- archived chats;
- files;
- the Memory Summary;
- connected apps containing the information.
That is because current Memory can synthesize context from more than one source.
If your goal is privacy rather than troubleshooting, review Settings > Personalization > Memory Summary > Manage and the original sources containing the information.
Turning Memory off prevents normal Memory use while disabled, but it is different from deleting every underlying source.
Memory Seems to Work on the Web but Not the App
First verify that both devices are signed into the same account and workspace.
Then update the official ChatGPT app.
OpenAI’s 2026 Memory improvements have required current app versions as features have rolled out to mobile.
After updating:
- Fully close ChatGPT.
- Reopen it.
- Confirm Memory is enabled under Personalization.
- Start a normal chat.
- Ask about a harmless detail you already confirmed is remembered on the web.
If the same account and workspace behave differently after an app update, the difference may be app rollout or interface state rather than lost memory.
Reinstalling should be a later step, not the first one.
The Memory Controls Are Missing
A missing control is an availability problem, not proof that stored context was corrupted.
Check:
- the active account;
- the active workspace;
- whether the app is current;
- ChatGPT on the web;
- administrator restrictions in a managed workspace.
OpenAI continues to evolve Memory controls and interfaces, so older screenshots and tutorials may show labels that no longer match the current product.
Use the options currently displayed under Settings > Personalization as the starting point.
Check OpenAI Status Only When the Problem Is Broad
An OpenAI service incident is worth checking when Memory behavior changes suddenly across multiple chats or devices.
Use the official OpenAI Status page when:
- personalization suddenly disappears everywhere;
- several ChatGPT features are failing at the same time;
- the same account behaves abnormally on both web and mobile;
- settings that were available suddenly fail to load.
A single forgotten detail is not strong evidence of an outage.
If other ChatGPT features are also malfunctioning, the comparison becomes more useful. For Voice-specific problems, see How to Fix ChatGPT Voice Mode Not Working.
Do Not Delete Everything to “Reset” Memory
Deleting all chats and memory data is a poor troubleshooting step.
It destroys the information you are trying to test and can make the account less personalized afterward.
Before removing anything, establish whether the problem is:
- Memory being disabled;
- Temporary Chat;
- an incomplete Memory Summary;
- one detail not being retained;
- one response not using a relevant memory;
- the wrong account or workspace;
- Project-specific behavior;
- a broader ChatGPT issue.
Only delete information when deletion itself is your goal.
When Memory Is Actually Working Normally
Memory is probably functioning when ChatGPT can carry a harmless preference into a separate normal chat, even if that preference is not visible as a standalone line in Memory Summary.
The summary is not a database, and Memory is not a verbatim archive.
That is the key distinction behind many reports of “ChatGPT Memory not working”: the expected behavior is often different from how the current Memory system is designed.
Use normal chats for the cross-chat test, Memory Summary for a high-level review, Sources to understand personalization when available, and Custom Instructions when an explicit instruction needs to remain consistently visible. Those checks tell you much more than repeatedly asking ChatGPT to remember the same detail.
