These points may already be familiar to experienced ChatGPT users.
However, I think they are very important for ordinary users who begin using Projects or Memory with the expectation that they can safely support long-term work.
I am not claiming to have discovered undocumented internal behavior.
My point is simpler:
OpenAI should explain these three boundaries more clearly, because the current product descriptions can easily lead users to expect more continuity and completeness than is actually guaranteed.
Related feature requests have already asked for better Project retrieval transparency and full-history search. My request here is slightly different: even before such features are added, I think the current limitations should be stated clearly enough that ordinary users can make informed decisions about what they can safely rely on.
1. A long conversation should not be treated as a complete, continuously preserved memory
Even if a long conversation remains visible in the ChatGPT interface, users should not assume that every detail from the entire conversation will remain equally available and usable indefinitely.
OpenAI’s Memory documentation already makes clear that Memory does not retain every detail from every conversation, and that ChatGPT determines which available information is relevant to a response.
The problem is not only simple forgetting.
For long-term work, a more dangerous failure can occur when the main rule is retained but its scope or exception is lost.
For example:
- “This rule applies only in this conversation.”
- “This applies only under these conditions.”
- “There is an exception in this specific case.”
If the rule remains but the limiting condition disappears, ChatGPT may produce a very plausible but incorrect answer.
For important long-term work, users should therefore not assume that something stated once will always be preserved and applied later with exactly the same meaning.
2. “A Project can reference other chats” does not mean “the entire Project is exhaustively read”
I think this is especially easy to misunderstand.
The high-level description of Projects presents them as a way to bring related chats, files, and instructions together so that ChatGPT can work with continuing context.
However, the more detailed documentation uses language such as:
- ChatGPT can reference other conversations in the same Project.
- For Plus and Pro users, ChatGPT prioritizes chats and files from the Project.
These statements are not necessarily contradictory.
But the following are very different guarantees:
The system can reference another chat.
The system searches every chat in the Project for every answer.
The system retrieves the full text of every relevant chat.
The system exhaustively reviews all Project material without omissions.
A normal user can very reasonably interpret “Project memory” or continuing Project context to mean that information placed inside the Project will reliably be found when it becomes relevant.
If that level of completeness is not guaranteed, I think this limitation should be stated much more prominently in the Projects documentation and interface.
The issue is not whether cross-chat context exists.
The issue is that:
“can reference” ≠ “searches everything” ≠ “reads everything” ≠ “exhaustively reviews everything.”
3. Naming a past chat and asking ChatGPT to “read the entire conversation” does not by itself guarantee a complete review
This is also very important for long-term work.
Here I am specifically referring to a past conversation being accessed through ChatGPT’s conversation history, Memory, or Project context — not to a case where the user has separately supplied the complete source text as a file or other explicit input.
A user may say:
“Please read this past chat.”
“Please review everything written in that conversation.”
“Please audit the entire conversation.”
But naming a past conversation and requesting a complete review should not automatically be interpreted as proof that its complete text was retrieved from beginning to end and exhaustively checked without omission.
These are different states:
The chat was referenced.
Relevant information was retrieved from the chat.
The full text of the chat was retrieved.
The entire text was reviewed exhaustively.
For tasks where completeness matters, that distinction is critical.
If ChatGPT cannot actually perform a complete review of the requested source, it would be much safer if the system explicitly said something like:
“I was not able to verify the entire conversation in full for this answer.”
That would be much better than allowing the user to assume that a complete review occurred.
What I am asking OpenAI to do
I am not asking OpenAI to pretend that unlimited perfect memory or exhaustive retrieval is always technically practical.
I understand that long-running conversations, retrieval, context limits, and computational cost create real constraints.
That is exactly why I think the boundaries should be clearly disclosed.
Users should be able to understand, in ordinary language:
What ChatGPT can access.
What it actually searched or referenced.
What it did not check.
Whether a review was complete or partial.
For Projects in particular, even a simple clarification such as:
“Project chats may be referenced when relevant, but the system does not guarantee that every Project chat will be searched or fully reviewed for every response.”
would prevent a great deal of misunderstanding.
In the future, it would also be very useful if ChatGPT could show:
- which previous chats or files were used for the current answer,
- whether the retrieval was partial or exhaustive,
- and whether a requested full review could actually be completed.
A simple way to check this yourself
If you are unsure about any of these points, you do not need to rely on my interpretation.
Ask your own ChatGPT to check OpenAI’s current official Help documentation first, and then ask it:
- Do you continuously retain the complete text of a long conversation and have equal access to every detail later?
- Does a Project guarantee that every chat in the Project is searched in full for every answer?
- If I name a specific past chat and ask you to review the entire thing, can you guarantee that its full text was retrieved and exhaustively checked without omission?
The important distinction is between:
being able to reference something
and
having completely retrieved and exhaustively reviewed it.
For experienced users, this may be obvious.
For users who are beginning to rely on ChatGPT for long-term Projects, research, writing, or other important work, I think it should be made explicit.