Recall + novelty

Catch the thing
you never wrote a rule for.

You don’t give it a list of bad things to watch for. You let it experience what normal is like. After that, anything new gets pulled toward the closest thing it remembers, and how far it has to travel tells you how well it fits.

How it learns normal, how it flags the odd one out, and what it unlocks.

01 · Learn normal

Show it what normal looks like.

There’s no rules table here, and no list of things to forbid. You simply let the memory experience ordinary inputs: a quiet office, a healthy machine, a typical transaction, and it remembers those examples as they come. That’s the whole setup. The memory now holds a picture of what usual feels like.

Writing the rules for “what counts as wrong” is the part that’s impossible everywhere else: you can’t list every bad case in advance. So you don’t. You teach it good, and let everything else give itself away.

It learns what your normal spending looks like from a handful of everyday charges, and from then on it can flag the charge that doesn’t fit. No labels, no examples of fraud.

02 · Pull to the nearest

New input gets pulled toward what it knows.

Hand it a new input and it doesn’t look anything up. It pulls the input toward the closest thing it has stored, the nearest memory it has. If the input is ordinary, that pull is short: it lands right on top of something familiar.

The distance it had to travel is the answer. A small distance means “I’ve seen this kind of thing,” a good fit. That number comes out on its own, without any threshold you had to hand-pick for every case in advance.

03 · The odd one out

What it’s never seen stands on its own.

Something genuinely new sits far from everything the memory holds. The pull is long and the fit is poor, so it stands out on its own, without anyone having described it first. Recognition by resemblance, not a rules table you had to keep up to date.

And because the memory keeps taking in ordinary inputs, it keeps learning what normal is. When normal drifts, whether a busier room, a season changing, or a machine wearing in, the picture moves with it. Nothing to retrain, nothing to re-deploy.

15 charges

of everyday spending to learn the shape of your normal

examples of fraud needed; it only ever sees your normal

live

keeps learning as normal drifts: no retrain, no redeploy

Try it yourself: teach it a handful of everyday charges so it learns your normal, then craft a new one. An overseas $4,200 charge at 3am gets flagged, because it sits far from everything it just learned.

04 · What you can do with it

Teach it normal. Let it flag the rest.

Experience

Feed it ordinary inputs and it remembers them as they arrive. No labelling, no list of forbidden cases, no training run. The picture of normal builds itself from what actually happens.

Measure the pull

Every new input gets pulled to the nearest thing it remembers. The distance it travels is the score, a fit, not a verdict. The familiar lands close; the strange lands far.

Surface the odd one

What it has never seen sits apart from everything it holds, so it stands out on its own. Recognition by resemblance, and the memory keeps learning, so normal can move without breaking it.

05 · What it unlocks

Catch what no rule could name.

Once you can teach normal and let everything else give itself away, a whole class of things a memory that thinks can do opens up.

Fraud you didn’t predict

Catch the transaction that looks unlike anything legitimate, without a fraud team writing a rule for every new scam first.

Machines about to fail

Learn the hum of a healthy motor; flag the new vibration that nobody had a name for yet, before it breaks.

Security that adapts

Spot the login, the process, or the packet that doesn’t resemble normal behaviour, even an attack pattern never seen before.

Quality on the line

Teach it a good part; let every off one give itself away, with no defect catalogue to maintain as products change.

Health out of the ordinary

A reading far from a patient’s own normal stands out on its own. It's personalised, because the baseline is theirs, not a population average.

Drift you can live with

As normal shifts over time, the memory keeps learning it. Alerts stay meaningful without a quarterly retrain.

What it powers

Notice what doesn’t belong.

This is the mechanism behind the Anomaly demo. It learns what your normal spending looks like from a handful of charges, then flags the one that doesn’t fit.

Open the live demo →Things a memory that thinks can do Download HeatherDB