Long-running agents do not only need recall. They need a durable state model that tells the next run which route survived, which branch failed, and which evidence must stay auditable.
The problem is state, not storage
Most memory systems answer a retrieval question: which text looks relevant? Long-horizon agents need a harder answer: which remembered state is allowed to influence action right now?
Aionis records the work loop as execution evidence. Plans, decisions, tool outcomes, verifier results, and handoff boundaries are admitted into explicit surfaces before the next prompt is compiled.
The execution contract
The runtime compiles memory into use_now, inspect_before_use, do_not_use, and rehydrate. That contract gives the agent a current route, keeps risky memory visible but gated, and preserves raw evidence without flooding the prompt.
This is the core difference between execution memory and ordinary recall: the system is deciding memory authority, not just memory similarity.