August 24, 2026 · Joy · What JoyMux guarantees · Part 3
What JoyMux keeps (and what the agent sees)
The agent does not have to remember what happened. JoyMux knows.
That split is the product: reasoning stays in the model. Execution evidence stays on the machine — a durable record you can replay, plus a bounded view you can hand to the next prompt.
Guarantee
Keep the complete execution record locally while giving the next agent only the relevant change.
Dumping full scrollback into the next prompt feels safe. It is usually waste: tokens, attention, and a laptop holding duplicated histories for every harness.
JoyMux keeps the record. The agent gets a bounded context view — what changed, plus enough surrounding state to act. Keep all the evidence. Send the AI only what it needs.
That is send the delta, not a style preference.
Replay is a read, not a reconstruction
Guarantee: execution evidence remains available independently of the model’s memory.
If you missed output, you do not reconstruct it from a vanished buffer. You replay from exact evidence.
JoyMux persists execution evidence before making it available for durable replay. Slow clients catch up from a known cursor. They do not stop the process from producing output.
See replay missed output for the Operator view of the same idea.
What this is not
JoyMux is not compressing your job into a slogan. The ~10× token figure from founder testing is secondary and said honestly. The headline is still: less rediscovery, less lost state, less “paste the whole terminal just in case.”
Don’t trust the write-up. Test the invariant.