The features we ship turned off
A memory layer stuffed with clever mechanisms sounds better than one that runs only what it can prove. Curator ships the clever ideas disabled when measurement refutes them, treats a revert as data, and keeps live only the parts that earned their place.
Adding machinery to a memory system is easy. One more ranking signal, one more reasoning pass, one more heuristic that sounds obviously correct the moment you describe it. The hard part, the part almost nobody does, is turning those things back off when the measurement says they did not help.
Curator does that, on purpose, and it is one of the reasons the recall stays sharp as the system grows.
Sounds right is not the same as measured right
A fair number of our own best ideas died on their first honest run. A mechanism that should plainly improve recall gets built, gets benchmarked against the same store without it, and comes back flat or negative. When that happens, the mechanism does not get quietly shipped anyway on the strength of how reasonable it sounded. It ships disabled. It is there in the product, off in your session, because it did not clear the bar.
That feels strange the first time. You wrote the thing, it was smart, and you are shipping it turned off. But a clever mechanism that does not move the number is not neutral. It is one more moving part that can misfire, one more way to be confidently wrong, added for a benefit that measurement could not find. Off is the correct place for it until that changes.
A revert is data, not a failure
The same holds after something is already live. When a change starts costing recall, it gets reverted, and the reason gets written down next to it: what we expected, what actually happened, and what would have to be true for the idea to work later. That record is worth as much as a win. It stops the same dead end from being wandered into again by a future session that only remembers the idea sounded good.
What you get out of this is a short list. The mechanisms running against your memory are the ones that each earned their keep on measured evidence, not a drawer full of plausible ones quietly dragging on quality.
Off today is not off forever
None of this is deletion. A signal that paid nothing last month can decide everything once a layer beneath it is repaired, so disabled ideas are re-tested, not thrown away. We hold a hard line against ripping out a component just to make an instrument happy, because the cheapest way to move a number is often to break the architecture that will matter next month. Fix how the pieces compose, re-measure, and let the evidence decide again.
Why fewer live parts is the point
The intuition that a memory layer is better because it has more mechanisms is exactly backwards. Every part that does not measurably help is a liability you cannot see: latency, a failure mode, one more way for the wrong thing to surface with confidence. Shipping the unproven ones turned off is not caution for its own sake. It is what a memory you can actually trust looks like from the inside.
Where it stands
Curator is v0.3.0, a working eval-gated pre-release product dogfooded daily across the projects that build it. The measurement discipline behind this is laid out in a memory you can measure, and the reason a leaner store beats a fuller one is in why AI memory gets worse. It answers to the same proof standard as BenchProctor and pairs with TheAuditor, Warden, and Arbiter. Follow release preparation.
Was this useful?