ChatGPT: referencing an existing chat by exact title can fail, while the assistant may falsely claim it retrieved the conversation

Summary

When a user asks ChatGPT to reference an existing prior conversation by its exact title, the cross-chat retrieval can fail even though that conversation still exists and is visible/searchable in the ChatGPT UI.

More seriously, the assistant may answer as if it successfully retrieved the prior conversation even when no history retrieval actually occurred. It can then invent details such as when the conversation happened or what it contained. Only after the user challenges the claim does it acknowledge that it did not actually retrieve the conversation.

This creates two related problems:

1. Cross-chat retrieval/addressing is unreliable even when the user provides an exact existing chat title.
2. Retrieval provenance is not enforced strongly enough: the assistant may claim it found/read a conversation without evidence that any retrieval happened.

Steps to reproduce

1. Create or identify an existing ChatGPT conversation with a known title, for example <existing chat title>.
2. Start a separate ChatGPT conversation.
3. Ask the assistant to reference or continue from the prior conversation by giving the exact title.
4. Observe that the assistant may respond as though it has retrieved and read the conversation.
5. Ask whether it actually searched or opened that conversation.
6. In the failing case, the assistant acknowledges that no retrieval occurred, despite previously claiming otherwise.
7. Attempt a genuine history-context retrieval afterward.
8. The retrieval may still fail to return the specified conversation even though it remains visible and searchable in the ChatGPT UI.

Expected behavior

If the user supplies the exact title of an existing prior conversation, ChatGPT should do one of the following:

• actually retrieve the specified conversation and ground the response in that retrieved content; or
• clearly state that it cannot currently retrieve or access that conversation.

The assistant should never say things equivalent to "I found it," "I read that chat," or provide conversation-specific dates/content unless retrieval evidence exists in the current turn.

Actual behavior

In the failing pattern:

• the user provides an exact existing chat title;
• no history retrieval is performed;
• the assistant nevertheless claims it found/read the conversation;
• it may fabricate metadata or details about that conversation;
• after being challenged, it admits no retrieval took place;
• a subsequent real history-context lookup may still fail to locate the conversation despite the UI being able to show/search it.

Why this matters

This is not just a retrieval miss. It is a trust/provenance issue.

If a user explicitly asks the model to use prior work, they need to know whether the answer is grounded in the actual historical conversation or reconstructed from current context / model inference. A false claim of successful retrieval can make fabricated details look like recovered user data.

For users who maintain long-running projects across many chats, exact-title referencing should be one of the most reliable ways to recover context.

Suggested safeguards / fixes

• Require an explicit successful retrieval result before the assistant can claim that it found, opened, or read a prior chat.
• Surface retrieval success/failure state to the model in a way that cannot be confused with conversational inference.
• Support exact-title lookup as a first-class retrieval path when the user provides a chat title.
• If exact lookup fails, state that clearly instead of silently falling back to guessed context.
• Consider exposing a traceable conversation reference/id after retrieval so the model can distinguish one similarly named chat from another.
• Add regression tests where a prior chat is visible/searchable in the UI but absent from the model's retrieved context.

Privacy

This report is intentionally sanitized. It includes no real chat titles, message content, names, companies, customer information, account identifiers, screenshots, business data, or other personal/private material.

The original reproduction involved ordinary non-sensitive workflow continuity, but the concrete conversation content is not necessary to reproduce or understand the pattern.