Sync agent state by shipping only the surprise.
An agent's KV-cache, memory, and context are dominated by frozen and appended history — successive states are near-identical. at1 statesync sends the byte-diff from the receiver's prior state and reconstructs the new one byte-for-byte, SHA-verified.
- 9–37×
- fewer bytes than re-sending full state
- byte-exact
- SHA-256-verified reconstruction, or it refuses
- format-agnostic
- works on any serialized state — no tensor assumptions
- grows with runtime
- the longer the agent runs, the bigger the win
KV-cache
Causal attention never rewrites the past — each step appends one block. The delta is that block; the frozen history rides free.
Long-term memory
A running embedding / memory store that mostly appends with a sprinkle of in-place edits — the diff catches both.
Context / scratchpad
An append-only message log plus a small mutable scratchpad — structured, byte-level, still an order of magnitude smaller.
Two commands
# sender: compute the delta from the receiver's prior state at1 statesync diff prev_state.bin cur_state.bin -o delta.bin # receiver: apply the delta -> byte-exact new state (SHA-verified, refuses on mismatch) at1 statesync apply prev_state.bin delta.bin -o cur_state.bin
Honest scope
The win comes from agent state being append-dominated — frozen KV rows, growing logs, mostly-append memory. If your state is regenerated wholesale each step (shuffled, re-embedded end-to-end), the delta approaches the full size and you should just send it. Cold-start (no prior) sends the full state once. Reconstruction is always byte-exact or it refuses.