In the first post I mentioned the moment the memory system was born: I was tired of starting from zero every time. This post is what happens when you take that idea seriously and point it at research itself.

Here is a confession that should embarrass me more than it does. For the first stretch of running this, we kept answering the same questions over and over. Not identical questions, but close enough that the work was mostly redone. A topic comes up in March. We dig in, produce something good, act on it. In May a related question lands, and the agent starts from a blank page, as if March had never happened.

Every research effort has this problem. Most never notice it, because the default shape of research is write-once, read-never. You produce an answer, you use it, and it goes into a folder nobody opens again. The work was real. The asset it should have become never formed. Most AI use has no memory at all: it answers, forgets, and pays full price to learn the same thing again next time.

The fix sounds too obvious to bother writing down: before you research something new, check whether you already know it. In late April 2026 I made that a hard rule, the first thing that happens on any new question, before a single new search runs.

It is a small step that does a lot. When a question comes in, the system first searches our own accumulated knowledge and reports back how much of the answer we already hold. Sometimes the verdict is “we have basically answered this; here is the existing work, go cite it.” Sometimes it is “we know the edges but not the middle, so only research the middle.” And sometimes it is “we have nothing, go do the full dig.” Three honest outcomes: skip, narrow, or go.

What that one step does to the economics is the whole point. Without it, every question costs full price forever, and the library just sits there getting bigger and more useless. With it, every question you have ever answered quietly lowers the price of the next related one. Knowledge stops being a cost you paid once and becomes an asset that pays back a little every time you reach for it.

I caught myself once about to launch a small fleet at a question, the expensive option, only for that first check to come back and tell me we had already covered most of it in earlier work. The honest move was to refresh the few stale bits and skip the rest. That check saved the whole effort.

There is a discipline hiding in here that is easy to miss. The check only pays off if you actually trust it. The temptation, every single time, is to run the full research anyway, just to be thorough. That instinct feels responsible and is actually the trap. If your own library cannot be the first place you look, you do not have a library. You have an archive, and an archive is just a graveyard.

And the loop only compounds if you close it. Every answer has to flow back into the same library the next question will check against. Research that never returns to the well lowers no future price. So the end of an investigation is not the deliverable. It is the moment the deliverable becomes searchable for next time. Most people stop one step too early, at the answer. The asset gets built in the step after.

So we search less and reuse more, and the work gets cheaper as it piles up. But cheaper and faster surfaced a different danger, the one that took me longest to respect. When an agent hands you an answer and a green checkmark, how do you actually know it is true? That is next.