Persistence became a benchmark. Here is ours.

Something quietly important happened in embodied-AI research this summer: persistence became measurable. A wave of new benchmarks scores persistent household world-state, memory consistency across long horizons, and whether an agent's beliefs survive contact with a changing environment. The field has started measuring what a robot believes, not just what it completes. We have been arguing that continuity is the missing axis of robot evaluation; it is good company to be in.

The bar is still too low for a home

But a benchmark built for a single agent in a single body misses what a household actually needs. A family does not need a robot that remembers well. It needs a home whose memory survives the robot. So here is the test we hold our own system to, in three clauses, each harder than the last.

One: memory survives a body swap. Serialize the mind on one robot, load it on a structurally different one, and the world model, preferences, permissions, and task state must arrive intact, with nothing body-specific leaking through. Two: authority survives the swap too. An action the household never permitted must be refused on the new body exactly as it was on the old one, with a machine-readable reason, not a silent failure. Three: every recall and every action is attested. The record of what the mind knew, decided, and did must be signed, tamper-evident, and replayable, so that a claim about the robot's behavior can be verified rather than trusted.

Runnable today, not aspirational

These are not goals. They are tests, and the v0.1 runtime passes them in simulation today: a humanoid learns a home, its mind transfers to a fixed arm that has never seen the room, the arm completes the task with navigation scoped out by contract, and every event on both bodies lands in one verifiable chain. The code, the 40-test suite, and the uncut demo output are public in our repository, cereal-os on GitHub, precisely so this claim can be checked rather than believed.

An RFC, not a victory lap

We are publishing this as a request for comment. If you build embodied-memory systems, benchmarks, or the motor intelligence our layer sits above, we want to know where this bar is wrong: too easy where homes are hard, too hard where it does not matter. And we hold the honest register on the field's behalf: nobody, including us, has shipped the uncut, unassisted version of the hard household tasks. Persistence is the axis where we intend to be measured first. hello@cereal.bot.