Published May 18, 2026
Today I activated the first piece of v3, the autonomous side of DARA. Since this morning the system tends to itself: it reads what's new, finds connections between memories, cleans up dead references, and writes its own one-line summary when a memory is missing one. There is no human in the loop and no queue of suggestions waiting for review.
The first day wasn't clean.
Five hours after launch, the system had drawn five wrong connections between unrelated memories. It saw shared category words ("website", "marketing", "design") and treated them as evidence of the same topic. That made five false positives across as many memories.
I stopped it and made it stricter. I raised the threshold for what counts as overlap and added a list of generic category words it should ignore. I also rewrote the underlying judgment so it demands a concrete shared subject (same project name, same person, same product) before suggesting any link. Then I ran a smoke test and relaunched it.
By evening, the autonomous side was running cleanly on a schedule. It starts when I sign in and restarts itself if it crashes. The fixes held, with zero false candidates after the iteration.
This is how v3 is supposed to work. With each step the system gets a bit more autonomous: you find where its judgment falls short, you correct it, and you let it run again.
As of today the housekeeping is delegated. Eyes and ears come next. Before adding them, I want a few more days of clean runs to confirm that the stricter linking rules keep false connections at zero.
Javier
A memory that tends itself, one iteration at a time. Early access is open.