@trq212 I second this! I've faced issue with markdown rendering in my native IDE. I've always had to upload the document to a site that has a proper rendering engine. So, I switched to getting HTML documents.
I was thinking of the exact same problem today, and arrived at the same conclusion. I suspected the current approaches to solve the memory won't scale well. Storage could be solve but retrieval is the problem. The precision & recall for RAG & text search degrades at scale. And, yeah, if the model doesn't know what to retrieve or if the stored memory can't be retrieved it's a dead memory.
Taking the human brain as analogy, I think we embed behavioral things directly into our weights (continuous fine-tuning) & store book knowledge in some representation (similar to text / embeddings) with a decay. If the learnt knowledge isn't utilized, it's slowly disappears from the memory. Likewise, when the knowledge it frequently retrieved, it eventually get's into the weights of our brain & body (i.e. muscle memory).
After using claude code for while now, making a switch to codex, cursor or any other app is quite hard since I've customized it with the skills, MCPs & hooks, etc. I wish I could take my setup with me when I switch.
@mattpocockuk You should probably add Alibaba's OpenSandbox as a backend. It'll allow organizations to run sandbox in their choice of infra & isolation.