As a user on both web/desktop app (chat interface) and Codex, I would be benefiting from a personalized experiences if memory management between ChatGPT and Codex can be unified, persists cross different devices / platform.
The other is ‘memory’ portability. As someone who uses multiple LLMs, I would greatly benefit from memory portability to any LLM endpoints I want to point to without sounding like a broken records 
Welcome to the forum!
Take this search for a spin:
Memory MCP
No sense in reinventing the wheel. 
Thanks for pointing me to the ‘Memory MCP’ thread previously shared by @teamlayer . It was an interesting idea and may partially solve the use case I mentioned here. Albeit a bit ‘hacky’.
My ask is for a native solution by OpenAI for Codex. Knowing what to keep in short (in session) and long (retrieval) is non trivial task and quite complex; attempted myself last winter with Google ADK ‘memory bank’ and compaction features.
Fast forward 6 months, with new techniques like ‘Skill’, I can see some of the memory management task can be offloaded to skill.md, particularly procedural ‘memory’. For semantic memory, potentially, can leverage the agent.md/LLM.md. But there still no solution for episodic memory. And if anything, that is what makes you, the user, ‘You’.
I actually was thinking you would use Google, forgot all about the @teamlayer topic, nice to know.
With stats like these hard to remember it all.
While I do not have crystal ball I do know that keeping tabs on the Codex commits, can foreshadow much of what OpenAI is doing in some areas.
e.g. memory
Why are you acting surprised 
JK aside, I think a custom built Memory MCP might be good for prototyping - to understand what is possible and constraints, but not for production even for personal use. On-going management and audibility is hell.
Hey @flamedagger, really appreciate this breakdown. You’re getting into the part of “memory” that’s actually hard 😅
Skills/skill.md helping with procedural memory makes a lot of sense, and agreed that episodic memory still feels like the missing piece. The portability point across ChatGPT, Codex, devices, and even other LLMs also comes up pretty often from power users.
Also appreciate @EricGT pointing out the Memory MCP.
Definitely forwarding this to the team for logging. Super thoughtful feedback here.
-Mark G.
Appreciate the acknowledgement
