Week 0x00

I guess it’s time to start deploying the household agents. Shannon and I decided a while ago to call them “bears” after a discussion about the man or bear meme.

I spent too much time thinking about this over the last few months. I’ve suspected for a long time that we needed to focus on the agent layer rather than fine-tuning an open-weights model, but only when I read the MemGPT paper did I really get that memory could be properly decoupled from model knowledge enough that we could build up something of long-term value.

That led me to the MemGPT team’s explainer and I decided to deploy Letta.

This is also an opportunity to vibe after learning to trust Kilo Code with some scoped admin tools for Hexmap.

So I’ve started with some infrastructure-as-code. (Or do we call it “gitops” now?) LiteLLM and Letta are up and running.