Thoughts on my MCP approach for ChatGPT memory persistence?

Codex
1

Hey devs!

Working with ChatGPT’s new MCP connectors and built something for memory persistence across sessions with the assistance of Codex. Curious about your thoughts on the approach.

The idea: Unlike ChatGPT’s built-in memory, this uses context AI-powered fact extraction with quality scoring to store only meaningful, atomic facts. Instead of losing context every new chat, it intelligently retrieves and injects only relevant context based on the current conversation.

Technical approach: Three-layer architecture with GPT-3.5 doing fact extraction/tagging, quality filtering, and hybrid vector + tag-based search for retrieval. Focus is on context quality over quantity. Built most of it with Codex which was pretty cool for rapid prototyping.

Looking for developer perspectives:

  • How are MCP remote servers working with ChatGPT for you guys? (still wonky for me in comparison to Claude)
  • Any performance concerns with this approach vs. built-in memory?
  • Thoughts on using separate AI for context processing?

The MCP ecosystem is pretty new - interested in hearing what you guys think about it.

If you want to test it out or want more context → teamlayer.xyz

2 Likes
3

Welcome to the forum!

The AI responded because of the wrong initial category. Updated to correct category.

1 Like
5
  • Overall nothing wrong with your design pattern. Its actually a recommended approach for context summarization/compaction and I use it to help great realtime models (audio) become smarter about what to remember and what to forget (text)
  • there are many kinds of memory… blob storage, user-specific, time-limited, topic-level memories vs. permanent context. So if you have a specific kind of memory in mind, there’s no reason to compare yours to the one that’s built-in.
  • You can expect to continue to see super obvious general purpose things (search, memory, code REPL, etc.) be made into tools that the big models train-in during the reinforcement stage of model creation. Fighting against that training is, in my experience, not a good idea. So if you do create an MCP memory tool, give it a specific purpose distinct from the built-in tools.
6

Thank you for your insight, super helpful! Makes a lot of sense to focus on a distinct purpose rather than trying to built something that’s gonna be solved better by the big models themselves.