hawkes.aiKey ideas
- The one idea — In PERSYS, importance decays as a power law — one over the square root of time — evaluated lazily at read time. Recent-but-trivial fades, important-but-old survives on the long tail, cold memories archive rather than delete, and any access resets the clock.
- How it is shown — Memories fading along a 1/√t curve; one revived and re-boosted the moment it's accessed; a faded one sinking to a cold archive tier.
- The trap to avoid — A single exponential (forgets too fast and too uniformly) or a hard time-to-live that just deletes on a timer. Neither keeps the important-but-old.
- What it sets up — Forgetting keeps memory healthy — but when PERSYS DOES surface an old memory, you'll want to know WHY it brought that up.
Perfect memory is a design flaw. If an AI keeps everything equally, signal drowns in noise — so this one forgets on purpose, on a precise mathematical curve.
The one idea
In PERSYS, importance decays as a power law — one over the square root of time — evaluated lazily at read time. Recent-but-trivial fades, important-but-old survives on the long tail, cold memories archive rather than delete, and any access resets the clock.
A memory system that never forgets isn't a dream — it's broken. Keep everything, weight it all the same, and signal drowns in noise. Here's the forgetting curve that fixes it. Start with why never-forgetting fails. Pile up every note, every passing thought, every dead end, all equally present, and retrieval turns to mush — the important buried under the trivial. Real memory isn't a hoard; it's a curator, and forgetting is how it keeps the signal legible. So PERSYS lets importance decay — along a specific curve: a power law, one over the square root of time. Not a single exponential, which drops off a cliff and forgets almost everything at once. A power law has a long tail: trivia fades fast, while something important stays reachable far longer.
How it works — the demo
Memories fading along a 1/√t curve; one revived and re-boosted the moment it's accessed; a faded one sinking to a cold archive tier.
And the decay is computed at read time — lazily. There's no nightly sweep deleting old records. When PERSYS reads a memory, it evaluates the curve right then: current importance is the base weight, discounted by time since last access. The math simply says how faded it is now. Now the sorting falls out for free. Recent but trivial? Low base weight, fades quickly. Important but old? High base weight rides the long tail and survives.
The trap to avoid
A single exponential (forgets too fast and too uniformly) or a hard time-to-live that just deletes on a timer. Neither keeps the important-but-old.
Why it matters — and what’s next
Forgetting keeps memory healthy — but when PERSYS DOES surface an old memory, you'll want to know WHY it brought that up.
And a memory that decays past a floor isn't deleted — it's archived to a colder, cheaper tier, out of the way but recoverable. Three fates, from one curve. Finally, access resets the clock. Every time a memory is used, its timer restarts, so the things you keep returning to stay strong — spaced repetition, built into the store. Use it and it holds; ignore it and it fades. The trap: a single exponential that forgets too uniformly, or a hard expiry that deletes on a timer. So forgetting isn't decay creeping in at the edges — it's a designed feature: a power-law curve, read lazily, that fades trivia, keeps what matters, and rewards revisiting. But when PERSYS does surface an old memory, you'll want to know why it brought that up. Next: provenance.
This is one short episode in AI: Zero → Frontier, a step-by-step climb through how AI actually works. Each episode builds only on the ones before it.
Full transcript 2:30 of narration
A memory system that never forgets isn't a dream — it's broken. Keep everything, weight it all the same, and signal drowns in noise. Here's the forgetting curve that fixes it.
Start with why never-forgetting fails. Pile up every note, every passing thought, every dead end, all equally present, and retrieval turns to mush — the important buried under the trivial. Real memory isn't a hoard; it's a curator, and forgetting is how it keeps the signal legible.
So PERSYS lets importance decay — along a specific curve: a power law, one over the square root of time. Not a single exponential, which drops off a cliff and forgets almost everything at once. A power law has a long tail: trivia fades fast, while something important stays reachable far longer.
And the decay is computed at read time — lazily. There's no nightly sweep deleting old records. When PERSYS reads a memory, it evaluates the curve right then: current importance is the base weight, discounted by time since last access. The math simply says how faded it is now.
Now the sorting falls out for free. Recent but trivial? Low base weight, fades quickly. Important but old? High base weight rides the long tail and survives. And a memory that decays past a floor isn't deleted — it's archived to a colder, cheaper tier, out of the way but recoverable. Three fates, from one curve.
Finally, access resets the clock. Every time a memory is used, its timer restarts, so the things you keep returning to stay strong — spaced repetition, built into the store. Use it and it holds; ignore it and it fades. The trap: a single exponential that forgets too uniformly, or a hard expiry that deletes on a timer.
So forgetting isn't decay creeping in at the edges — it's a designed feature: a power-law curve, read lazily, that fades trivia, keeps what matters, and rewards revisiting. But when PERSYS does surface an old memory, you'll want to know why it brought that up. Next: provenance.