TLDR:

I wrote a Codex Plugin for Forgetful Memory MCP, which has my opinionated workflow for managing context across agents and sessions.

---

Hello folks,

I hope this post finds you well, as well as it does me, as I've finished a nice weekend grind outside of my day job working on my hobby, which is anything to do with AI Agent CLI's.

I bloody love'em and especially tinkering with them and getting a feel for what the latest harenesses and models are capable of.

I spend most of the week reading about all the cool stuff other people are doing with them, and then at the weekend

At the heart of a lot of this tinkering sits my own memory framework I built some time ago, Forgetful Github Repo. I won't spend too long discussing it here, suffice to say - it is an MCP service that is highly configurable and provides tools to expose a Graph-RAG service for persisting agent memories and forces agents to adopt a Zettlekasten approach to storing and retreiving knowledge.

Since probably March of last year, I have been a Claude Code user for the most part, with some usage of copilot (enforced at work), Gemini CLI and Opencode (local models) here and there.

I built a plugin around Forgetful on Claude Code when they introduced them and built a nice little workflow with Forgetful and Context7 (great tool, the first thing I ever recommend to a fellow developer who is picking up agentic coding tools for the first time).

Since then I and others have contributed some skills and commands to the repo for other harnesses, but I had never used Codex much myself and nor has any submitted a PR for anything beyond the basic setup.

The Codex Plugin

Recently with the free month offer I have picked up Codex and decided build a Forgetful plugin with it as a nice little way to see how it compares to the other agentic harneses that I have used.

What I currently have with other implementations, is very much a manual memory management workflow. I use slash commands (or skills if you prefer) to invoke when the agent should go get information and likewise curate information in forgetful's knowledge base.

To this end the plugin contains some skills for memory curation, context gathering (including the use of context7 if it is available), project initiation and encoding repos (taking a repository, adding it as a project in forgetful and then making atomic notes about it).

I find this encode-repo approach to be really powerful when I work across complex architecture involving many components and repositories. When an agent queries forgetful, it is not is not looking to retrieve the code, it is looking to retrieve what it can remember about the code.

Encode-Repo Skill

In the same way a human developer might do so when they are asked a questions about a product, they will recall which libary or repository the logic related to the question sits and knows to look at the code for more specific information if required.

My normal work flow when starting some work is to ask an agent to use the context-gather skill and then give it a prompt or a topic, it will then launch sub agents to bring back relevant context from forgetful, context7, code repostories and also the web.

Context-Gather Skill

I then build a plan with the agent with that context. Once implemented we then finish by asking the agent to use the memory-curate skill, which is basically an instruction for it to ensure that the knowledge base now refelects any of the changes implemented as a result of the plan, including the removal of anything stale.

This feels right for me where I am, I have used this workflow with Claude since Sonnet 4.0.

This is why as of now there are no hooks or other automated memory curation, I personally have not found a workflow that I feel strongly enough about to say to everyone else using Forgetful "This is how you should automate memory curation".

I tried some initial hooks but they more often than not bloated the CLI screen, context window and usage. So I've left them out for now. That is not to say there is not a way to do this, many others have appeared to have settled on it and look to be extremley popular. It's just I, myself, have not!

GPT 5.5 to Opus 4.7 Comparison

Anyhow, on a slightly different note I must say I have noticed a big difference in performance from GPT 5.5 compared to Opus 4.7.
To the point I am contemplating switching over as it generally feels like Anthropic has really fumbled since Feb with its frontier model. It's fine for my own workflow where I pretty much hand hold it through each step, but if I want to do anything a bit more hands off (like some cool side project) it just really struggles.

I had been working on some plugins for my Hermes agent with Opus 4.7 on the side and it just kept getting it wrong, GPT 5.5 has just come in and basically one shotted each request after I started them from scratch. Purely ancedotal of course, I have nothing to baseline either against.

Ironically another of my own side projects is my own set of evaluations, so eventually I'll have something more concrete than ancedotal/gut feel impression of this, but figured I'd throw this opinion out there while I am here.

Anyhow I guess that wraps it up for this wall of text. Thanks for reading if you made it this far and if you like the plugin and make use of it great and any PR's/issues are welcome on Github.

Happy coding