Feature request: Compact long ChatGPT conversations into a clean new thread

This is not just branching.

Branching forks from a specific point in time.
Compacting preserves the current source of truth from the whole conversation while discarding obsolete context.

The problem is not that users need another chat.
The problem is that long-running chats accumulate stale context, and the model may keep paying attention to decisions that have already been reversed.

Additional use cases where this would be especially valuable:

  1. Hallucination correction / invalidated context cleanup

In long chats, the assistant may make an incorrect claim, the user corrects it, and the conversation continues. However, the original incorrect answer still remains in the thread and can continue to pollute context.

A compaction feature should not simply summarize everything that was discussed. It should distinguish between:

  • active facts and decisions
  • superseded decisions
  • invalidated assistant claims / hallucinations
  • user corrections that should become the new source of truth

Invalidated claims should not be carried forward as usable context. The compacted thread should preserve the corrected fact and explicitly exclude the hallucinated answer.

This is different from ordinary summarization. The goal is not to preserve the full conversation history. The goal is to preserve only what remains valid.

  1. Travel planning

Travel planning conversations often change direction many times:

  • different cities are considered and rejected
  • dates change
  • budget changes
  • hotels or flights become unavailable
  • the user changes priorities
  • earlier recommendations may no longer apply

A compacted new chat could carry forward only the final itinerary, confirmed constraints, remaining open decisions, and rejected options with reasons. This would be much cleaner than continuing a long thread full of outdated travel assumptions.

  1. Coding / debugging

In coding sessions, many hypotheses are explored and later disproven. The assistant may suggest a cause, the user tests it, and it turns out to be wrong. Later, the real cause is found.

A clean compacted thread should preserve:

  • the confirmed root cause
  • the current patch direction
  • files already changed
  • tests already run
  • known failed approaches
  • remaining tasks

It should not continue to treat disproven debugging hypotheses as relevant context.

  1. Research and analysis

In research-heavy chats, the user may compare many sources, reject weak evidence, correct factual mistakes, and refine the thesis over time. A normal branch does not solve this, because useful conclusions may appear near the end while invalid assumptions may exist in the middle.

Compaction would let the user create a clean research state:

  • verified facts
  • unresolved claims
  • discarded sources or weak arguments
  • current thesis
  • open questions
  • next research steps
  1. Writing / planning / product work

For long writing, strategy, or product planning conversations, decisions often evolve. Earlier outlines, names, requirements, or priorities may become obsolete. Continuing the same thread can cause the model to keep referencing old directions.

A compacted thread could preserve the current source of truth:

  • final direction
  • accepted structure
  • active requirements
  • discarded alternatives
  • tone/style decisions
  • next draft actions

The main issue is not conversation length alone. The issue is stale or invalid context surviving after the user has already moved on.

A useful mental model:

Branching preserves the timeline.
Compaction should preserve validated context.

Suggested behavior:

  • create a new chat from the current conversation
  • generate an editable handoff summary first
  • separate active context, superseded context, and invalidated context
  • carry over relevant files/artifacts where possible
  • make the compacted summary the new thread’s source of truth
  • keep a link back to the original conversation for provenance

This would help long-running ChatGPT workflows stay accurate after decisions change, assumptions are rejected, or hallucinations are corrected.

Hey @gagul, that’s a valid ask, and it sounds like it covers a different use case than the existing branching feature.

I’ll make sure the feedback is surfaced to the engineering team so they can evaluate whether this is something worth implementing.

Avinash

One important requirement: relevant files and artifacts should be carried over as well.

Long-running ChatGPT workflows are often not just text conversations. They may include uploaded PDFs, screenshots, spreadsheets, code files, generated documents, images, canvases, or intermediate outputs.

If a compacted new chat only carries over the summary but loses the relevant files/artifacts, the user has to manually re-upload or reconstruct the working context, which weakens the value of the feature.

Suggested behavior:

  • identify files/artifacts still relevant to the active source of truth
  • let the user choose which ones to carry over
  • exclude files tied only to discarded or invalidated directions
  • preserve file names and provenance
  • optionally include a note explaining why each file was carried over
  • keep a link back to the original conversation

The ideal flow would be:

Compact into new chat
→ generate editable handoff summary
→ show active / superseded / invalidated context
→ show relevant files and artifacts to carry over
→ user confirms
→ create clean new thread with both the summary and selected files attached

This matters because the new thread should not just preserve a text summary. It should preserve the usable working state.

Existing compaction in AI tools is usually optimized for token/context management: summarize older context, keep critical state, discard less relevant details.

The requested feature is slightly different: validity-aware compaction.

It should distinguish:

  • active facts and decisions
  • superseded decisions
  • disproven hypotheses
  • invalidated assistant claims / hallucinations
  • user corrections that should become the new source of truth

This is especially important because ordinary summaries can accidentally preserve hallucinated claims as “things discussed.” A clean handoff should not preserve everything discussed; it should preserve only what remains valid.

Appreciate the additional context on validity-compaction.

- Avinash