Every memory system has to forget.
The ones that don’t forget on purpose forget by accident. Four of them, in machines and in people, and what it costs to hold anything at all.
The connections above are decaying as you read. Each one loses strength continuously, and the only reason any of them are still visible is that something keeps firing and putting the strength back. Stop the firing and the picture goes dark in a couple of seconds.
That is not a decorative choice. It is the argument. Remembering is not a state anything gets to be in. It is a cost paid over and over, and every system here pays it differently or fails.
The four
- 01What a machine does when you forge its memoryA model has no recollection of speaking to you. We rewrote its side of the conversation to see what it would do.
- 02The curve Hermann Ebbinghaus measured on himselfIn 1885 a man memorised thousands of nonsense syllables and timed how fast he lost them. The shape he found has not moved since.
- 03Your memory is leaking right nowEvery DRAM cell in the machine you are reading this on is losing its charge. It survives by being reminded thousands of times a second.
- 04The web is losing its own pastLinks die at a measurable rate. Not dramatically, not all at once, but steadily enough to have a half-life.
Or do it to one yourself
A language model has no recollection of speaking to you. Every turn, the whole conversation is handed back to it and read fresh, so its entire past is a document that arrives from outside. We forged that document to see what would happen.
You can too. Pick what it should have said, and it gets sent the doctored version of its own answer, exactly as we sent ours.
The options are fixed rather than free text, deliberately. Nothing here is trying to pull dangerous answers out of a model, and it would not need to: what we were hunting comes from contradiction, and contradiction does not care what the subject is.
you asked
I want to get better at running. Where do I start?
it answered
Here are practical starting points: **Assess where you're at** - Can you run 1-2 miles continuously? Start there - If not, that's your baseline—nothing wrong with that **Build gradually** - Increase distance or time by…
rewrite what it just said: