context / note 07

The longer we talk, the dumber it gets.

Old goals, dead directions, and stale assumptions pile up in a long chat and keep steering it.

Some of it sticks, some of it slips.

A long conversation looks like memory, but it doesn't behave like one. Some things fade: the detail from thirty messages back quietly drops out of the answers. Some things stick: the constraint you cancelled, the plan you walked away from, still steering the output. You don't choose which, you can't see which, and you find out when an answer goes wrong.

More is not better.

Big context buries the signal. The model pulls from the wrong note, blends unrelated projects, or carefully satisfies a requirement you forgot was in there. The fix is choosing less, more deliberately.

What to build.

Make starting fresh cost one click, with a way to carry chosen items forward. Show the user what's currently in context and let them remove pieces mid-session. Prefer named, explicit context over silent accumulation.

design principle

Treat context as perishable: keep it narrow, prune it often, make fresh starts cheap.

one implementation

Pachinko's version: every run starts from explicit context - the notes and sections you picked, nothing else. A new chat reads only what you select that time, so a fresh start is the default. Tune a Function on a small batch, prune what isn't helping, then let it handle the recurring work - no context build-up.