Feature request: Opt-in shared context between ChatGPT and Codex, including Memory and Reference Chat History

Codex
,
1

I use both ChatGPT and Codex under the same OpenAI account. My need is not only for Codex to access ChatGPT saved memories, but for ChatGPT and Codex to share the relevant context layer when I choose to enable it.

Currently, ChatGPT has two useful context mechanisms:

  1. Saved memories, such as long-term preferences, goals, coding style, research background, and personal workflow.
  2. Reference chat history, where ChatGPT can use useful information from past conversations to make future conversations more relevant.

For coding and long-term projects, Codex would benefit from both. Many of my coding tasks are first discussed and planned in ChatGPT, then implemented in Codex. However, Codex conversations are separate from ChatGPT conversations, so I have to repeatedly copy context, decisions, constraints, project goals, and personal preferences into Codex.

I suggest adding an opt-in “shared context” setting between ChatGPT and Codex.

Possible design:

  1. Off by default.
  2. User can enable “Allow Codex to use ChatGPT context.”
  3. User can choose the scope:
    • Saved memories only
    • Reference chat history summaries
    • Specific ChatGPT Projects only
    • Specific chats selected by the user
  4. Codex should show what type of context it is using.
  5. Users should be able to exclude sensitive memories or chats.
  6. Users should be able to revoke access at any time.
  7. For project work, a ChatGPT Project and a Codex workspace/repository could optionally share a project-level context summary.
  8. Codex should not need full raw chat history by default; it could use a user-approved summarized context layer to reduce privacy risk and noise.

This would make Codex much more useful for long-term software projects, research projects, and personal knowledge-management workflows, while preserving user control and privacy.

2

Fully agree this would be a sweet option to have shared context. And would keep them ahead of the competition who are not doing this either (yet). My guess (based on what has been announced by OAI and alluded to) is that we will see this in the coming weeks. Fingers crossed.

3

I would love it so much. I see several risks though:

  • Burning a lot of token cause the context on GPT can be huge and it would be loaded every query of codex
  • Risk of loosing the delays, I don’t know for you but codex for me has access to the real full code directly and so is clother to the truth on that end. On the other end GPT has the architecture of the project but can think that he knows the code. Fusing the 2 context can be easy for a human knowing the project but priorizing which context in each specific bit of the project would be extremely hard for either of the 2 AI. Unless you train a dedicated agent to whom you explain exactly how you are processing.

In the end I have to say I got used to it. It allow me to spot some non-sense that the two of them exchange with each other.

If the architecture is clear in your discussion with GPT and the goal that you want Codex to achieve with the prompt you ask GPT to make you’ll indeed still need to copy paste but it will work wonderfully.

Sometimes, when everyone only know what he needs to know it is also improving efficiency.

I do agree that a personalization area built in codex would be great. But you can work in project and explain it way less either.

4

I think that if they combined, but also kept the token use separate so ChatGPT work did not eat into Codex work tokens (as it is today) that would be awesome.