Multiple ChatGPT reliability failures affecting memory, document integrity, instructions, files, and conversation state

I am documenting a broader ChatGPT reliability problem affecting several parts of my paid ChatGPT Plus workflow. This is not a report about one isolated response, one long conversation, or a single interface error.

Between October 9 and October 21, 2026, I experienced recurring failures involving memory and continuity, document accuracy, instruction-following, information retrieval, conversation state, and access to accumulated work.

I have already reported these issues privately to OpenAI Support. I am posting here to determine whether other users have experienced the same combination of problems.

Environment

  • ChatGPT Plus
  • Windows 11
  • Google Chrome
  • Chrome Incognito
  • Opera
  • ChatGPT Windows desktop application
  • Timezone: Central Time, UTC-6

1. Memory and conversational continuity failures

ChatGPT repeatedly asked for information that had already been provided and discussed in detail.

This occurred across unrelated subjects, including ongoing projects, computer troubleshooting, purchases, personal organization, financial matters, and important documentation.

The recurring behavior included:

  • asking for previously supplied information again;
  • failing to apply saved context and established instructions;
  • reopening decisions that had already been made;
  • losing project hierarchy and confusing secondary tools with primary projects;
  • requesting completed material again;
  • failing to preserve corrections after acknowledging them;
  • recovering only fragments of prior information and transferring the remaining reconstruction work back to me.

The concern is not simply that ChatGPT cannot remember everything. In several cases, the system acknowledged that the information had already been provided, promised to preserve continuity, and then repeated the same pattern.

2. Errors in important documents

The most serious part of this case involves documents where mistakes can have consequences beyond an ordinary incorrect answer.

The problems included:

  • literal quotations attributed to people even though those quotations were not present in the original transcripts;
  • an institutional action being included as fact even though I had explicitly corrected that it had not occurred;
  • incompatible financial totals produced from the same underlying matter without a transparent reconciliation;
  • chronological corrections failing to persist;
  • interpretations being presented as verified facts;
  • documents described as reviewed or transferred for review without a completed deliverable or verifiable result.

I understand that model output must be verified. However, there is a material difference between an ordinary mistake and presenting invented wording inside quotation marks, inserting events that did not occur, or producing multiple incompatible calculations without identifying what changed.

3. Plausible reconstructions presented as established context

ChatGPT also converted possibilities, plans, or visual clues into facts.

Examples included:

  • treating an object I had once planned to build as if it already existed;
  • assuming that a tool visible in an older photograph was currently available;
  • connecting conditions from different areas of a room without adequate support;
  • interpreting residues as possible mold despite corrections about location and chemical use;
  • replacing a literal delivery notification with a simplified assumption, then treating my interpretation as the mistake.

These are not all memory failures in the narrow sense. They involve failure to distinguish:

  • verified fact;
  • previous plan;
  • current availability;
  • interpretation;
  • unverified hypothesis.

4. Explicit instructions and restrictions ignored

Clear requirements were also lost or overridden.

Examples included:

  • recommending products that violated a binary exclusion requirement;
  • disregarding previously established budget, size, comfort, and practical constraints;
  • changing visual elements that had been explicitly protected;
  • confusing the purpose or hierarchy of projects;
  • redirecting a clear request for a ChatGPT failure report toward an unrelated subject;
  • repeating troubleshooting or methods that had already failed.

Correcting one problem sometimes resulted in new changes to information that was already correct.

5. Library search and access to accumulated work

On October 17 and 18, a ChatGPT Library search returned some files and images but displayed:

“Some search results could not be loaded.”

The interface did not identify:

  • which results were missing;
  • how many results failed;
  • whether the visible results were complete.

I reproduced the same behavior in:

  • Google Chrome;
  • Chrome Incognito;
  • Opera;
  • the ChatGPT Windows desktop application.

I am not claiming that files were deleted. The problem is that the interface did not allow me to verify search completeness, indexing, accessibility, or preservation.

6. Interrupted processing and inconsistent client behavior

Other symptoms included:

  • a response ending with “Thinking interrupted” before completing a feasible task;
  • images or complete conversations displaying differently in the Windows application;
  • the same conversation presenting different states in Chrome, Opera, and the desktop application;
  • uncertainty about whether a requested review or background process had actually been performed.

7. Contradictory maximum-length warnings

On October 19, one conversation displayed the warning:

“You have reached the maximum length for this conversation, but you can keep talking by starting a new chat.”

The warning appeared at least three times.

However, the same conversation continued:

  • accepting messages;
  • accepting image attachments;
  • generating responses.

After another successful exchange, the warning appeared again.

The issue is not merely that a conversation limit may exist. The issue is that the interface stated that the limit had been reached while the same conversation remained operational.

8. Native Share feature failure

During the same period, attempting to use the native Share function produced:

“Unable to open chat sharing.”

This prevented me from creating a share link for the same conversation that was displaying contradictory maximum-length states.

9. Cumulative impact

These failures required repeated reconstruction of information, manual verification of documents, correction of invented details, rechecking of calculations, repetition of instructions, and troubleshooting across several clients.

The result was not one defective answer. It was an accumulated reliability problem affecting:

  • long-term continuity;
  • important documents;
  • financial reconciliation;
  • retrieval of stored work;
  • project development;
  • productivity;
  • trust in generated material.

Comparable public reports

I found other public reports describing related behavior:

  • saved memory being recalled but not reliably applied, followed by plausible context being presented as verified history;
  • loss of context between consecutive messages and custom instructions being ignored;
  • continuity errors and new mistakes introduced while correcting an earlier response;
  • existing conversations remaining visible but becoming inaccessible across several clients;
  • maximum-length warnings appearing in new or relatively short conversations;
  • a maximum-length warning and an “Unable to open chat sharing” error occurring in the same affected conversation.

These reports do not prove that every case has the same internal cause. They do show that the observable behavior is not limited to a single type of failure.

What I am asking the community

Has anyone else experienced a combination of the following?

  1. Saved information is available but not applied.
  2. Corrections fail to persist.
  3. ChatGPT reconstructs plausible context instead of verifying history.
  4. Quotes, facts, chronology, or document content are introduced without source support.
  5. Correcting one section changes unrelated information.
  6. Explicit restrictions or custom instructions are ignored.
  7. Library or conversation retrieval is incomplete or inconsistent.
  8. Conversation states differ between browser and desktop clients.
  9. Maximum-length warnings appear while the conversation remains usable.
  10. The Share interface fails on the affected thread.

I am not asserting a specific root cause. I am asking whether these behaviors are reproducible, whether they became more frequent after recent changes, and whether they are being investigated as related reliability, context, retrieval, or conversation-state issues.

One approach we’ve been working on to address conversation continuity is preserving the working state of a project: important decisions, constraints, context, and unresolved issues.

We developed a ChatGPT plugin called Continuity Relay that extracts this information into a portable XML file, which can then be used to continue working in a new AI conversation.

It doesn’t fix ChatGPT’s memory or reliability issues, and it’s not a verbatim conversation backup, but it can help preserve valuable work.

Continuity Relay is publicly available in ChatGPT’s plugin directory. You can search for it by name and try it for free, without purchasing anything.

I’m one of the developers, and I’d genuinely appreciate feedback on whether this approach helps with the continuity issues you’re experiencing.

Hey @Mayte, thanks for documenting this, and thanks @felixeduardog for sharing your suggestions.

Are these issues still happening, and has Support been able to help? If they’re still unresolved, please follow up on your existing Support request or share your existing case number here and I’ll review what’s already been reported.

- Moizz

Hi Moizz,

Thank you for responding.

Yes, these issues remain unresolved, and additional problems have occurred since my original post on October 22.

My existing OpenAI Support case is #17504159.

I have already reported multiple incidents through Support, but I have not received a substantive resolution. The problems have continued to affect memory and conversation continuity, document integrity, file accessibility, and reliability across ChatGPT interfaces.

More recently, I have also experienced unexplained Work usage attribution, problems processing valid PDF files, and additional application errors.

I want to emphasize that these failures have had consequences beyond ordinary inconvenience. I rely on ChatGPT Plus to organize, document, preserve, and retrieve important personal information accumulated over time.

The repeated failures have caused substantial additional work, delays, and difficulties recovering information that I had already documented. Some of these consequences are time-sensitive, and repeating or reconstructing the same work does not resolve the underlying problem.

My next subscription renewal is scheduled for November 14, while these issues remain unresolved.

Since you offered to review what has already been reported, I would appreciate a review of case #17504159 and clarification on whether these recurring failures can be investigated together rather than treated as separate, unrelated incidents.

I am not looking to open another support ticket or restart the reporting process. I am asking for meaningful follow-through on the case and incidents that are already documented.

Please clarify what the next step will be following your review of my existing case.

Thank you.
Mayte

Hey @Mayte, thanks for sharing your case number. I’ll review what’s already been reported and follow up with you through your existing private Support case when I have an update. I’m closing this topic so we can continue supporting you there.

- Moizz