How AI Works
$definesession--plain-english
TLDROne continuous conversation with your AI.
I used to treat one chat like my only shot. Same thread for days — scaffold the stack, paste errors, argue about CORS, rewrite the todo page — because starting over felt like throwing away progress.
It was not progress. It was clutter I refused to put down.
Think of a session like a work shift. Everything between clock-in and clock-out lives in that one shift. The next person who walks in knows nothing about it — unless somebody wrote it down and handed it over.
That is a session with an agent. Your first message. Every reply. The files you dropped. The tangent about Polar vs Stripe. All of it lives inside one stretch of chat, inside one context window the model re-reads every turn.
Close the tab. Open a new chat. A fresh worker clocks in with a blank clipboard.
The slate wipes every time. That sounds like a bug. It is the feature.
When a long chat gets confused — forgetting the stack, looping on the same fix, contradicting itself — your instinct is to push harder. Re-explain. Correct it again. Do not. You are not broken. Neither is the model. The shift is buried under three hours of noise, and a new shift starts clean.
The move takes thirty seconds:
You keep the gold. You leave the trash. The new worker reads a clean page instead of digging through the mess.
Compaction is the host trying to summarize the shift while you stay in the same room. Prompt caching is prep that survives between messages — as long as you do not edit the prefix. A new session is the honest version: close the door, open another, hand over the note.
Same word shows up on the web after auth: a cookie that means "this browser is still that user." Different badge. Same idea — a stretch of time where the other side is allowed to remember you.
One conversation. One session. One shift. Close it and the slate wipes. Once you know that, you stop being afraid to start over. Starting over is the feature.
