Published June 16, 2026
Last week the system contributed a fact: a supplier’s bank account number, pulled from a real conversation. The curator processed it, opened the right note, and found that the file already listed a different number for the same supplier — from a different source, on a different day, at similar confidence.
Six months ago, my system would have just kept the new one. Three months ago, it would have kept both and confused the next person who looked. Two months ago, it would have flagged it for me to resolve by hand. Today it flagged it for the next AI that touches the memory.
The autonomous curator runs around the clock. When it can’t decide between two contradicting facts, it doesn’t pick — it writes the conflict to a queue. That queue is part of the memory’s standard read protocol: any AI that reads the memory’s context also gets the queue, under a header that says “these need a decision.” The next AI to connect — me in a Claude session, an outside assistant over the protocol — sees them on the way in.
Here’s the part I find interesting: I built four ways for those flags to close, and only one of them involves me.
One — by consistency. If a new fact arrives on the same subject as an open flag and the curator finds it non-contradictory, the flag closes itself, logged as resolved by newer data. No human needed.
Two — by another AI. Any AI talking to the memory can resolve a flag: it passes the flag’s ID, an explanation, and supporting evidence. The flag closes with the resolution logged forever — which AI, which model, which session, what evidence.
Three — re-opening by another AI. If a resolution turns out wrong, any AI within twenty-four hours can re-open the flag, with the rejection logged. The window is deliberate: it forces disagreement to be timely, not a permanent litigation channel.
Four — by symmetry. Two clean cycles after an infrastructure warning disappears, the flag closes itself. Same principle as consistency, applied to the plumbing instead of the facts.
No single mechanism is the point. The property that emerges from all four is this: the queue stays small without me. New conflicts arrive, old ones close, and the shape of the flow looks like a real human escalation queue — except the actors are autonomous and the audit trail is structured.
I ran a smoke test on it. I told one AI to resolve a flag with a made-up justification. It closed. I told a second AI the first was wrong; it re-opened, rejection logged. A real cycle later, the underlying fact had stabilized and the flag closed by consistency — for real this time. All without me clicking anything.
What stays mine is the strategic call: “we should restructure the routing,” “this note is becoming a dumping ground,” “the summaries are drifting from their sources — we need to keep them in sync.” Those still come to me. The point of the four mechanisms is to make sure they’re the only thing that does. The operational back-and-forth — facts contradicting, resolutions, re-openings, warnings going stale — handles itself.
The pattern I keep finding: as the autonomous side gets more capable, my own surface area shrinks. Two weeks ago the memory needed me to grade it. Last week it started grading itself. This week the conflicts between AIs working on the same memory resolve themselves. Each version leaves a smaller mailbox waiting when I sit down. The mailbox isn’t empty — it may never be. But the things that reach it are the ones that actually need me. That’s the only thing that ever mattered.
— Javier
EIDARA v2 is free. SUPER DARA is what comes next.