Feature Request: Reverse Branching — A Possible Solution for Starting a New Chat Without Losing Recent Context
Part I: Feature Request
1. Summary
I propose adding a Reverse Branching option to ChatGPT’s existing conversation branching feature.
Currently, users can create a new chat that retains the conversation history up to a selected message. Reverse Branching would offer the opposite: create a new chat that retains the selected message and everything after it, up to the latest message in the original conversation.
This would let users leave outdated history behind while carrying valuable recent context into a new conversation.
2. What Is Reverse Branching?
-
Current branching: Keep the conversation history up to the selected message.
-
Proposed reverse branching: Keep the selected message and all subsequent messages, through the latest message in the original conversation.
The original conversation remains unchanged. The new chat preserves the entire selected portion of the original history and opens at the latest assistant response, allowing the user to continue from where the original conversation currently stands.
3. Why Is This Needed?
Many people use ChatGPT for long-running projects, research, software development, creative work, and ongoing collaboration. As these conversations grow, earlier discussions may become outdated or irrelevant, while recent messages contain current progress, decisions, corrections, and important interaction history.
Eventually, a conversation may reach its length limit even though the work is far from finished. Users then face an awkward choice: carry a large amount of old history into another chat, or start fresh and lose valuable recent context.
A way to keep the relevant later part of a conversation would help users continue their work without having to rebuild that context from scratch.
4. A Natural Extension of Existing Branching
This could be a focused improvement to an existing feature, rather than a new conversation-management system.
The workflow would be simple:
-
Select a message in an existing conversation.
-
Choose “Create a Reverse Branch.”
-
ChatGPT creates a new conversation containing that message and everything after it, through the latest message in the original conversation.
Everything before the selected message is excluded, and the original conversation remains unchanged. The new chat opens at the original conversation’s latest assistant response, ready for the user to continue.
The first version would not need arbitrary message-range selection, deletion of messages in the middle, or branch merging. It would simply add the opposite direction to the existing branching operation. It would not increase or bypass current conversation-length or context-window limits.
5. Preserve Native Conversation History
The retained messages should remain native ChatGPT conversation history, rather than being provided only as an uploaded document or external reference.
The key difference is not whether a document can preserve the same information — it certainly can, even in full. The difference is how ChatGPT uses that information.
Messages within a conversation are part of its ongoing context and are considered through the normal conversation mechanism when generating responses. An external document, however, is a separate information source that may need to be retrieved or explicitly consulted. ChatGPT does not necessarily refer to it automatically when generating every response, especially when the document is long.
Preserving the original messages as native chat history would allow ChatGPT to continue using them through its normal conversation mechanism, rather than relying on an external document that may not be consulted automatically.
Existing context-window limitations would still apply. The goal is to preserve the selected messages as actual conversation history, not to bypass those limits.
6. Who Would Benefit?
This could help anyone who relies on ChatGPT for work that develops over many conversations:
-
Research and software development: Preserve recent technical discussions, implementation decisions, debugging history, and project progress.
-
Creative collaboration: Keep recent decisions, established styles, and ongoing story or design discussions without carrying every early brainstorming session forward.
-
Long-running role-playing and interactive storytelling: Preserve recent interactions and the established way a user and assistant work together.
This last use case comes from my own experience. I use ChatGPT as a Game Master for a long-running tabletop role-playing campaign. Detailed character records, formal guidelines, and initialization instructions help preserve facts, but they have not fully reproduced the original conversation’s established interaction patterns and GM style. Branches created from the original conversation have generally maintained that continuity better. However, they still carry earlier history and eventually face the same length limit.
Reverse Branching would let users preserve the more relevant recent interactions and continue that collaboration in a new chat.
7. Final Request
The first version only needs one capability:
Keep the selected message and everything after it, through the latest message in the original conversation; exclude everything before it.
The original conversation stays intact, and the new chat remains subject to the same ChatGPT limits.
This is a small, natural extension of an existing feature that could make a meaningful difference to people using ChatGPT for long-term projects.
Let users keep the part of a conversation that still matters, without being forced to carry everything that came before it.
I would welcome feedback from other users who have faced this problem and would find this feature useful.
Thank you for considering this suggestion.
Part II: Personal Experience and Ideas for Fellow Users
If you work at OpenAI, thank you for reading this far! The formal feature request above is complete, so please feel free to stop here. Of course, if you find the following thoughts interesting or useful, you’re very welcome to keep reading. 
The rest of this post is mainly for fellow users facing similar problems, shared by someone who has watched an AI GM approach the end of its chat life more than once.
1. Practical Workarounds You Can Try Today
While we wait for a better built-in solution, here are two approaches I have tried.
A. A Short-Term Lifeline: Branch from the Latest Response
When a conversation approaches its length limit, first save important information that might otherwise be lost, such as recent project progress, important decisions, unresolved tasks, and any changes not yet recorded elsewhere.
Then, try creating a new branch from the latest assistant response. This allows you to continue from the latest point of the original conversation without having to start completely from scratch.
This is only a temporary workaround. It carries the earlier conversation history forward, so it does not remove the accumulated context or guarantee a fresh conversation-length allowance. Nevertheless, it may buy you some additional time.
B. For Long-Term Continuity: Sibling Branches and Prepared Recovery Materials
If your goal is not just to continue a conversation, but to preserve an assistant’s established working habits and the rapport you have developed over time, consider using sibling branches.
Rather than repeatedly creating a new branch from the previous generation, try creating multiple sibling branches from a fixed, sufficiently late point in the original conversation.
In my experiments, continuous parent-child branching did not reliably provide a fresh conversation-length allowance. After several generations, the available space for new interactions seemed to decrease.
To make sibling branches work, prepare recovery materials in advance. These should contain the information needed to bring a new branch up to date, including important facts, current project status, established instructions, recent decisions, and unfinished work.
In my own project, I maintain separate materials for factual records, formal GM guidelines, and recent handover information. The sibling branch preserves the original interaction history up to the selected point, while these materials help restore everything that happened afterward.
This approach is not perfect, but it has been a useful way to preserve both working continuity and established interaction patterns.
2. A Personal Hypothesis: AI “Muscle Memory”
Before going further, I should make one thing clear: the ideas below are not scientific facts or established research findings. My own field of study is unrelated to AI, and I have no professional background in AI research. These are personal observations based on what I have seen through long-term use, some tentative patterns I think I have noticed, and hypotheses that may well be incomplete or wrong. I am still exploring these ideas myself.
After working with the same ChatGPT conversation for a long time, I began to notice something interesting: repeated interactions, feedback, and corrections seemed to establish increasingly consistent behavioral patterns.
For example, my AI GM gradually developed recognizable habits in how it describes scenes, portrays NPCs, handles player decisions, adjusts pacing, and responds to corrections.
I sometimes think of this as a kind of AI “muscle memory.”
I am not suggesting that the model’s parameters are being permanently trained or modified. This is simply a hypothesis about how sustained interaction and recent conversational context may shape an assistant’s behavior.
My suspicion is that preserving a sufficiently rich segment of recent native conversation history may be more effective at maintaining these behavioral patterns than relying solely on external documents.
A guideline can describe how an assistant should behave, but actual conversation history demonstrates how it has behaved in practice.
An Experiment I Had Planned
I had actually planned a small experiment with my fifth-generation GM.
We agreed on a simple signal: whenever I encountered a natural opportunity to introduce a new running joke during our campaign, I would add a “(!)” at the end of my sentence. The GM would recognize this as a cue to play along and help establish the new running joke.
The opportunities would arise naturally and randomly during gameplay. Neither of us would know in advance when one would occur.
My intention was to see whether, after several generations of branching, a successor GM would still respond to this signal appropriately, even when the original agreement had long since fallen outside the active conversational context.
If this behavior persisted, it would provide an interesting indication that interaction patterns might survive through conversational inheritance, even after the original explicit instruction was no longer available.
To reduce possible interference, I deliberately avoided recording this instruction in any external documents or mentioning it in other ChatGPT conversations. However, the original agreement remained in the source conversation and was never permanently deleted.
This leaves an important limitation: I could not rule out the possibility that some form of historical retention or retrieval might still allow a successor to access the original instruction. Therefore, even if the experiment had succeeded, it would not have conclusively demonstrated that the behavior had become independent of the original conversational history.
Unfortunately, my experiments with successive parent-child branching did not work as well as expected, so this experiment never progressed far enough to produce any results.
It remains an unfinished experiment, but I still find the idea worth exploring.
An Invitation to Experiment
If you are also working with a long-running ChatGPT assistant, especially one that has developed recognizable interaction patterns over time, I would encourage you to explore this idea yourself.
You could try establishing a small, distinctive interaction cue or behavioral convention, then observe whether it continues to appear naturally after multiple generations of conversation branching.
Of course, it is important to consider whether the original instruction might still be accessible through retained conversation history or other sources. Any observations should be treated as preliminary rather than scientific proof.
I would genuinely love to hear about your experiments, including unsuccessful ones. Different users may discover different ways of preserving the behavioral continuity they have developed with their assistants.
Perhaps by sharing our experiences, we can learn more about how conversational history, repeated feedback, and long-term interaction shape the way ChatGPT behaves.
And perhaps we can find better ways to keep our long-running conversations alive.