A conversation that never ends — auto-compaction keeps decades-long threads alive on tiny context windows.
This one is still being built. It will open on a conversation that already has
years behind it: pre-seeded threads long past any model's window, kept alive
because context/compact folds the older stretches into summaries
the moment the working budget fills, without ever being asked. You pick up
mid-thread — suggested-reply chips give you somewhere to start — and the fold
happens under you, turn after turn, while the token meter refuses to climb.
Nothing is thrown away: any folded stretch can be opened back up to its full
original text when a question needs the detail a digest dropped.
Want to see hierarchical compaction working today? Chat with a
book folds an entire book into a small-context model and shows every API call it makes.
The code for both lives in
algolia/agent-studio-demos.