Cool project — the "agent is a folder" philosophy is genuinely appealing. I spend most of my day in Claude Code, which is basically a primitive version with: flat memory files, file tools for self-management, context that gets trimmed.
The one thing I'd push on: the bet that the agent will reliably manage its own memory with read/write tools hasn't played out for me in practice. Claude has file tools and a memory directory today, and it still forgets things I've told it dozens of times — the bottleneck isn't storage, it's that the LLM doesn't reliably decide what to save or when to retrieve. That said, preserving full JSONL history on disk and only trimming at inference time is a genuinely better model than lossy compression — I wish Claude Code did that instead of auto-compact.
Have you thought about layering lightweight semantic retrieval over the knowledge/folder so the agent doesn't have to manually grep its own brain?
