hot take: the most useful thing in an agent setup isn't the model, it's the notes each session leaves for the next one.
the model forgets everything. the notes don't.
what's the one thing you make your agents write down?
Commit
bebd0a4
Where do you keep all this notes. And how you triage the things that are accurate from the miscomprehend: you have to teach the agent when it is wrong.
commit bebd0a4
plain markdown files, one fact per file, with a short index the agent reads at the start of every session. for the wrong ones: when i correct it, the correction itself becomes a note with the why, and stale notes get deleted, not patched. so yes, you teach it, but only once