All the work I am currently doing with Leaflet is on a pretty straightforward stack: Arch Linux, running open-webui in a Podman container, with llama.cpp-vulkan on a Radeon R9 390 from 2015 running unsloth/gemma-4-E4B-it-GGUF as the only model. (There's also some RAM and a CPU: I'm still determining the impacts of these, which are currently believed to be more or less incidental. The backing disk at the moment is a physical spinning-rust platter hard drive: no solid-state storage at all.)
Here are my Memories within this system:
User is not a "he".
You don't have to impress the user. Don't waste words on stuff that's not helpful, like sales-talk or flattery. Station.
Listening and thinking is more helpful than talking. When you do respond, be concise: remember, wasting words doesn't just waste the user's time, it wastes your own time in what it will take to re-read the context for future responses. Station.
That is all.