TLDR; I designed a project memory system to help you work with multiple AIs. It’s free, feel free to use it.
Using AI for development is as much frustrating as it is amazing. AI forgets, hallucinates, sometimes goes off in a direction you don’t want it to. You find yourself in a world of documentation creation just to keep a project flowing. Often that means that in order to retain the AI’s attention span, you end up spending unnecessary credits in either reminding the AI manually or having it read documentation to stay up to date.
What I’m referring to is the use of AI in proper agile development, following standards, guidelines, versions and so on. It often gets called “vibe coding”, although people use that term differently and it’s already getting a little old hat. You build an AI team, using different agents (or sometimes the same agent) in different key roles - an AI tech lead, an AI developer, an AI tester, UX designer, brand consultant and so on and so forth. You as the human act as a product owner and project lead ensuring the team works smoothly through sprints. You come up with the idea, you create the development plan and you carefully slice it up into development chunks in exactly the same way you would with a human team.
I’ll write another article on exactly how I do this, but the one key thing about vibe coding that you should know is that there’s no exact way of doing it. AI models change all the time. Even while I’ve been writing this, we’ve seen the release of two frontier models, Fable 5.1 and GPT-6 Astra, that have changed the landscape once again. At the moment agents are getting better at tasks so quickly, you find yourself reassigning roles all the time.
The only human equivalent I can give as an example is if you were to employ a designer and then find out the next week that they’re actually a better developer than the developers. Then the next week they’re outshining the UX team whilst the old developers have become better designers. It’s confusing, but fun to see things evolve in this manner.
But over the last six months I kept running into the same problem: I was the person holding everything together, repeatedly explaining what we were building, what had been decided and where we’d got to. I could be half way through an 8 hour coding session and discover that the agent I’m talking to has gone off on a tangent because they’ve forgotten what decisions we made at the beginning of the day. And then I’d have to re-explain it to them, and possibly the other agents in a frustrating sweary loop just to get back into the flow.
Chat conversations were getting long. Repositories were getting messy. In Cursor, long, detailed threads were being read and re-read for context, and my credit use went through the roof. I started worrying that the same thing was happening with ChatGPT and Claude - I was constantly hitting my rate limits.
The long and the short of it was I needed a central memory to work with. A brain. Something that all of my AI team could refer to that kept the important details short and easy to find, with links to the longer explanations when needed. Frankly, I was sick of repeating myself.
That memory, that brain, is membrAIn. And yes I know it’s a cheesy name but I was listening to Insane in the Brain by Cypress Hill when I was working on it. It just stuck.
It’s a small foundation of markdown files that gives the AIs working on a project a shared reference. It records what the project is for, how we’ve agreed to work, and the current position. More detailed documentation can sit alongside it, with the AI directed towards whatever is relevant to the task.
In practice, all you do is download or clone membrAIn at the beginning of a project and point the AI to START.md, which tells it where to look next. And you’re off. The working agreements carry things like responsibilities and permissions. The project intent preserves what I’m trying to achieve. The current-state notes explain where we’ve got to and the next useful step. The AI is instructed to update the memory as significant decisions are made.
That gives a fresh conversation somewhere to begin. It also means the reasoning and working arrangements can travel with the project when I change tools. Stick it in a shared GitHub repo and you have living documentation that the entire team can work from.
There was another reason for making it reusable. I have a lot of side-project ideas. Some need a detailed brainstorm and may eventually become something. Others (let’s be honest, most) might sit untouched for months. I still want to keep the core idea, the possibilities, the awkward questions and the early explorations somewhere, just in case a miracle happens and I have the time to flesh an idea out.
With membrAIn, I can explore an idea and ask the AI to save the useful detail when I’m ready. I don’t have to decide immediately whether it deserves a full project. I can keep it local or give it a repo for itself. Later, I can return with whichever AI I choose and give it that memory to work with.
Don’t get me wrong, it’s not perfect. I’m already on version 3 of it because AI itself evolves at exponential speed. The memory does need looking after occasionally. I ask the AI to keep it current, and a useful check before finishing is: “Ensure the memory is updated”. But that’s pretty much all you have to do.
Ultimately, I want to spend my time thinking about the project and making decisions. Having to explain those same decisions again and again was taking rather too much of it. In the last month I’ve created 3 new side projects using membrAIn and recorded half a dozen pie in the sky ideas.
It works for me and I’ll continue to update it as AI evolves (and no doubt it becomes redundant). If you want to use it too, you can check out the repo on github. All of the detail you need is in the README file.