One memory + arithmetic

Add and subtract
the people you stored.

A person becomes one memory: the sum of everything they did. Similar people end up close together, so finding who’s alike is one step, not a pile of joins. And because each one is a sum, you can do arithmetic on them.

How a person becomes a sum, why that makes arithmetic possible, and what it unlocks.

01 · One memory

Everything they did, kept as one.

Take a person, or a product, a song, a customer, and everything they did. Instead of a row here and a row there, scattered across tables, the substrate folds it all into a single memory that is the sum of all of it. Every click, every play, every purchase, added into one thing.

That one memory is the person, in a form you can work with directly. Nothing is lost in the summing. What they did is still in there, and you can read it back out.

A user with two dozen actions is one memory built from all of them. Ask it what that person did and it hands the actions back, recovered from the sum, not fetched from two dozen separate rows.

02 · Similarity is one step

People who behave alike land close together.

When two people did similar things, their sums come out near each other. So “who behaves like this person?” is a single step, finding the nearby memories, not a query that joins a dozen tables and grinds through every event they share.

Lookalike audiences, related products, “customers like you”: the work that usually means a heavy pipeline becomes one short reach across memories that already sit where they belong.

03 · Arithmetic

Add for a blend. Subtract for a difference.

Because each person is a sum, you can do arithmetic with them. Add two people and you get a blend: recommend for a couple, build a persona that’s the middle of two real ones. Subtract one person from another and what’s left is the difference between them, a segment, on its own, with no description written in advance.

There is no operator for “this customer minus that one” in any database. You can’t SELECT it. Here it’s one line, and it disappears cold-start, too: the moment someone does a single thing, they have a memory, and the arithmetic includes them.

1 step

to find who behaves like a given person, not a pile of joins

a + b

two people blended into a third, to recommend for a couple

a − b

the difference between two people, read off as a segment

alice + dave − bob isn’t a query you could write in SQL. It’s arithmetic on memories, and what falls out is a group that nobody defined by hand. Live on the engine, updating as you change the inputs.

04 · What you can do with it

Sum it. Compare it. Do math on it.

Sum

Everything a person did becomes one memory. It doesn't grow into more tables as they do more. It stays one thing, and what they did can still be read back out of it.

Compare

People who behaved alike sit near each other, so similarity is a single reach, not a join across every shared event. Lookalikes and related items fall out for free.

Compute

Add two memories for a blend, subtract one for a difference. Segments and personas come out as arithmetic, things no database operator can express in a query.

05 · What it unlocks

Math on taste, not a pile of joins.

Once a person is a sum you can do arithmetic on, a whole class of things a memory that thinks can do opens up.

Recommend for two

Add two people and recommend for the blend, whether a couple, a household, or a team, instead of picking one and ignoring the other.

Segments by subtraction

Subtract one group from another and read off what makes them different. The segment defines itself; nobody writes its rules.

Lookalikes in one step

Find the audience that behaves like your best customers with a single reach, not a nightly pipeline grinding through events.

No cold start

The moment someone does one thing, they have a memory and they're in the math. No waiting for enough history to model them.

Personas you can move

Build a target as arithmetic, this taste plus a little of that minus the other, and shift it live by changing the inputs.

Taste, not a profile table

Identity lives as one composable memory, so the work that needed joins and feature pipelines becomes a line of arithmetic.

What it powers

Recommend a couple. Carve out a segment.

This is the mechanism behind two demos: Add and subtract what’s stored (analytics), and Blend two into a third (recommendation).

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

Add and subtractthe people you stored.