I wonder how scalable that is. After the twentieth feature has been added, how much connection will the conversation about the first feature still have with the current code? And you’ll need a larger and larger context for the LLM to grok the history; or you’ll have to have it rewrite it in shorter form, but that has the same failure modes why we can’t just have it maintain complete documentation (obviating the need to keep a history) in the first place.

Things like MemGPT/Letta, ToM-SWE, and Voltropy have made long context documentation pretty manageable. You could probably build some specialized tooling/prompts for development artifacts specifically too. But I’ll be the first to admit this is basically “Throw more agents at the problem”