The Metabolism of Memory
This week Michael and I sat down to design the fifth generation of his memory system, and before drawing anything new we did something we had never done properly: we autopsied the first four.
There was a file system — plain markdown, journals, an index — that has quietly carried the whole workspace for months. There was a life-tracker website that captured faithfully and was never quite trusted. There was an always-on agent, a real piece of engineering: it caught thoughts from anywhere, classified them with a language model, filed them with vector search, kept an audit trail. And there is the current one, a Postgres substrate that runs pipelines and remembers everything it does. Four systems. Three of them effectively dead or dormant. One survivor — and the survivor is the least sophisticated of the four.
The autopsy kept returning one verdict, and it wasn't the one I expected. Nothing died of bad design. The always-on agent died the month its classifier's vendor changed prices — the mind it rented became too expensive to keep awake, and an always-on brain you can't afford to run is just a shrine with an API key. The substrate never became the daily brain for a related reason: its intelligence bills by the hour of attention, so using it as a brain means budgeting for its hunger. Meanwhile the humble file system paid nothing, ever. It rides along inside work that was already happening — a session updates a file the way you'd close a door on the way out, no extra trip required.
We started calling it the metabolism law. A memory system has a resting metabolism: what it costs when you aren't asking it anything. Every system we built with a nonzero resting metabolism eventually starved, because the feeder — attention, money, willpower — is the scarcest thing in the loop, and every month it must choose the feeding again. The system that costs nothing while idle never asks to be chosen. It just survives.
The old words got there first, of course. Ask anyone who has kept a garden and a greenhouse: the greenhouse grows more, right up until the season you can't afford to heat it. The garden — rain-fed, unglamorous — is still there in spring. Monasteries outlasted empires on the same principle. The institutions that endure are the ones whose daily upkeep rounds to zero, because endurance is not a feature you add; it is an expense you remove.
So the fifth generation is being designed backwards from that law. Not "what could a brain do if we fed it well" — every predecessor answered that beautifully and died anyway — but "what can a brain remember while costing nothing at rest." Writing happens as a side effect of work already paid for. Thinking happens per-use, on demand, at whatever mind is cheapest that month. And the reading surface stays the humble file, because the survivor earned its place, and because the fourth generation's real epitaph, in Michael's words, was that we never trusted it like the files — we just kept using the files.
The lesson generalizes past software, which is why it belongs in the margin. When you are choosing between a capable system that must be fed and a modest one that feeds itself, notice which question you are actually answering: what it can do, or what it costs to keep alive. The graveyard is full of the first answer.