5d • General Discussion
I gave my codebase a memory that writes itself
I built Karpathy's LLM wiki but for my SaaS. After every Claude session it updates on its own.
My repo has a memory folder. Notes about my own product, written by Claude.
Every time I close a session, it reads all my conversations and auto updates the memory folder.
The whole thing is just hooks.
→ One runs when a session ends
→ One runs on every push
→ Both write to the folder
There's one on every prompt too, but that one only makes Claude tell me if what I'm asking for is even worth building.
No vector database.
No RAG pipeline.
No memory app.
Just markdown files sitting in the repo, committed to git like normal code.
And because it's in git, the memory isn't mine. It belongs to the repo.
Anyone who clones it gets the same thing. An AI that already knows the product, and a hook that saves their conversations into the same folder.
That's the part people miss.
Everyone builds memory for themselves. But everyone's conversations are sitting on their own machine. Now all of it ends up in one place.
So I don't explain my own product to Claude anymore.
It already knows.
Every day I get new thoughts about the product. Small improvements on what I thought yesterday.
I don't write any of it down. I just talk to Claude Code like I normally do, and the hook turns that session into memory.
So now I can see what I thought about my product on day one, and what I think today.
And how much it has changed in between.
If you want to try it:
Add a SessionEnd hook in your .claude/settings.json. Point it at the session transcript and tell it to append what mattered to a memory folder in your repo. Then commit that folder like normal code.
The part that matters is telling it to skip most sessions. Otherwise you just end up with a very long log.
0
0 comments
I gave my codebase a memory that writes itself
