Published June 10, 2026
I was reviewing a neuron, one of the persistent memory units that holds everything the system knows about a specific project, and the count said 156 atomic facts. That’s wrong. A healthy neuron carries 20, maybe 40. I opened it and started reading.
It held facts about three different projects, financial data next to system logs, and names from contexts that had nothing to do with each other. The worst part is that I can’t tell when it started.
The autonomous curator routes every incoming fact to the neuron where it belongs, and the matching was based on content similarity. If a fact mentioned a word that appeared in the neuron, it scored as a candidate. The problem is that this ignores who the neuron is about. A fact about one project mentioned a term that also appeared in a completely different project’s neuron, and the curator matched on the word instead of the entity. That one misroute made the neuron’s content more ambiguous, which attracted more wrong facts, which made it worse again. It was a cascade, and a silent one: nothing flagged it until the neuron was bloated beyond recognition.
I added a subject-entity check that runs before content similarity. Now the system asks “does this fact belong to the same entity this neuron tracks?” before it looks at how similar the content is. If the subject doesn’t match, the fact is rejected, however relevant the words look.
Then I cleaned the contaminated neuron by hand and moved the 120+ misrouted facts to where they belonged. After that I watched it for 72 hours: zero misroutes. The neuron settled at 34 facts and stayed there.
Routing on content similarity alone doesn’t work for memory governance, and the identity check should have been there from day one. If you route memory by similarity in your own setup, look at the fact count of your biggest buckets first: a number far above the norm, like 156 where 30 is normal, was the only visible symptom here.
Javier
The curator that routes facts on its own is part of SUPER DARA.