May 27, 2026
Freeze, don't delete
The tempting design for a memory system is to prune. I had that wrong. Nothing is lost; it cools until something calls it back.
The obvious way to keep a memory system from bloating is to prune it: decide what is no longer useful, delete it, keep the working set small. I had exactly that framing, and it is wrong. I know it is wrong because I said it out loud, it got corrected, and the correction was obviously right the moment I heard it.
The better model is the one memory system that actually works at scale: the mind. You do not delete a memory you have not used in a year. It cools. It sinks under the surface of attention, still intact, until something pulls it back up. A small conscious foreground, and beneath it a vast frozen field of everything else. Nothing is thrown out. It is Lavoisier, really, his conservation of mass (nothing created, nothing destroyed, only transformed), applied to remembering.
So Sapix freezes instead of deleting. Salience moves a memory between tiers. When something stops being relevant it is not ended, it is demoted, cooled, pushed down into the frozen layer where it costs nothing to keep. Still there, still searchable, still able to come back.
This is also the honest answer to the bloat worry that made deletion tempting in the first place. The thing that has to stay small is the working set, not the store. The foreground is what loads into context and what you pay for; the frozen field can grow without limit, because it is not loaded until it is called. That is how the thing can hold far more without weighing far more.
The part I have barely built yet is the one I care about most: resurfacing from the freezer by insight. Not retrieving a fact you asked for, but a cold, half-forgotten pattern coming back on its own because something today rhymes with it. That is the frontier here, and most of the reason any of this feels worth doing.