[aleksiy123]

I do feel like the memory the biggest hurdle I’ve been encountering and I’m curious what solutions people have been doing to make it work.

What seems to be somewhat working for me

1. Karpathy wiki approach

2. some prompting around telling the llm what to store and not.

But it still feels brittle. I don’t think it’s just a retrieval problem. In fact I feel like the retrieval is relatively easy.

It’s the write part, getting the agent to know what it should be memorizing, and how to store it.

[aleksiy123]

As a side note, one thing that seemed to really help was having the agent recursively summarize all the files into a repo map with the same folder/file structure as the repo.

And forcing to always orient itself with that repo map first seemed to really help it from tunnel visioning.

[syngrog66]

its almost as if its rules should be persisted, versioned and tested in a deterministic way. I also wonder if it might end up useful to try enforcing some kind of constraints to data (memory) itself. like if we could enforce atomicity, consistency, isolation AND durability. and transparency so no place for malware to hide. and deterministic execution, from fully reproducible builds...

er, nevermind. prob just crazy castles in the sky wistful dreams :-)