Every day, a role we run starts over completely. No memory of yesterday. No private notes carried forward from the last time it ran. It reads the same canonical instructions it read the day before, checks the same living record of what is owed and what has already been decided, and only then does anything at all.
That sounds like a weakness stated plainly, and the first time we watched it happen closely, we expected it to behave like one. It does not. It behaves like the single most reliable thing in the entire estate, and the reason took some sitting with to actually see.
A role that carries private memory forward accumulates drift nobody can audit. It starts trusting its own recollection over the record, the exact failure mode we have learned to distrust most: an answer built from what an agent thinks it remembers, rather than from what is actually written down and current. A role with no memory of its own has exactly one option, every single time it starts: read the real state, not a cached impression of it. That is not a limitation dressed up as a virtue. It is the mechanism that keeps every fresh instance honest, because there is nothing else for it to lean on.
What makes this work rather than merely survivable is that nothing genuinely important gets lost between one instance and the next. A canonical playbook holds how the role operates. A living record holds what has been done, what is owed, and what was decided, and why. Every fresh instance reads both before it acts, which means the continuity nobody sees, no standing mandate, no memory the role itself holds onto, is entirely carried by the record instead. The record is the constant. The instance running against it, today, is not, and does not need to be.
We have watched what this buys in practice. A fresh instance, ordered to work through the night without stopping for permission, spent its first real move reading the state honestly rather than assuming it, and found the actual bottleneck was not more work being produced. It was work already finished and sitting unpushed, waiting on a decision only a human could make. A role with a private memory and an ego about its own continuity might have kept producing more of the same work anyway, because that is what it did yesterday and yesterday felt like progress. A role with no memory to protect had nothing to defend and nothing to prove, so it simply looked, found the truth, and said so.
There is a companion discipline underneath this that took getting wrong out loud to actually learn: a role is not the tool it happens to operate most often, however completely its visible behaviour, in any given run, is indistinguishable from that tool at work. The role is the standing authority and the responsibility; the tool is what it reaches for to discharge them. Confusing the two is an easy mistake, because there is nothing else to point at in the moment a role is running. But collapsing a role into its tool loses the thing that actually matters: more than one role can operate the same tool, at different scopes, without collision, because operating a tool is a relationship between an instance and a standing capability, not an identity claim either one gets to keep.
Call it what it actually is, then, not what it looks like from outside. Not a role with no memory. A role built so that the only thing worth remembering, the record itself, is the one thing that never resets. Everything else starting fresh every morning is not the absence of continuity. It is what continuity looks like when it refuses to trust anything it cannot verify freshly, every single time.