Three Ways to Build a Memory

Three approaches. Two of them are what every tutorial teaches. Both decay.

Look at that last column. That's the lesson.

1. Meaning search

You'll hear this called vector search, or embeddings. Here's what it actually is.

The computer reads a note and turns it into a long list of numbers — a few hundred of them — that captures roughly what the note is about. Notes about similar things end up with similar numbers.

Then you ask a question. Your question gets turned into numbers the same way, and the computer finds the notes whose numbers are closest.

Why it's clever: it can match "how cold should insulin be" to a note that never uses the word "cold." The numbers for "cold" and "36 degrees" land near each other, because they mean related things.

That's genuinely useful, and it's why this is everywhere.

Why it decays: every note you add is one more thing competing for the top five slots. At 50 notes the winner is obviously right. At 5,000, the top five are all plausible and you can't tell from the outside which — if any — is the one you wanted.

Nothing degraded. The pile grew.

2. Linked notes

You'll hear this called a knowledge graph. Also simpler than it sounds.

Instead of guessing what's related, you write it down. This note is about that person. This appointment connects to that prescription. This bill belongs to that house.

Then you can walk the connections: start at "Ranger," step to "vet," step to "prescription food."

Why it's clever: it knows things meaning-search can't guess. Two notes with completely different words can be firmly connected because you said so.

Why it decays — and this one is sneakier.

As you connect things, a few items end up connected to almost everything. A person you mention constantly. A project. A year. Follow the connections from anywhere and you arrive at one of these hubs, and from the hub you can reach the entire pile in one more step.

Once everything is two hops from everything, "follow the connections" hands you back the whole memory. Which is the same as handing back nothing.

3. Both, plus cleanup

Use meaning search to find candidates. Use connections to pull in what relates. And then — the part almost nobody builds — let unused things fade, and delete what's genuinely dead.

The first two are how you find things. The third is how finding things stays possible.

This is what you'll build, and the reason the table's last column reads "no."

Why the fix has to be deletion

There's a temptation to solve this with better ranking. Smarter scoring. Cleverer search.

It doesn't work, and it's worth understanding why before you spend a month on it.

Ranking decides which of these 5,000 notes is best. That question gets harder as the number grows, no matter how good your ranking is, because more of them are genuinely close. You're asking a harder question every month and hoping for a better answer.

Deletion changes the question. Ranking 200 notes that have earned their place is an easier question than ranking 5,000, and it stays easier forever.

One approach fights the growth. The other removes it.

What this costs you

Being honest about the trade, since this is the whole design:

You will lose things. Not important things, if you build it properly — but the system will delete notes, and some of them will turn out to have mattered.

That's why everything you'll build keeps a receipt: every deletion is written down first, with the reason, so you can look at what went and put it back. A cleanup you cannot audit is a cleanup you cannot trust, and you would be right not to trust it.

It takes time to pay off. On a fresh install, a memory with cleanup and a memory without behave identically. You'll prove that yourself two lessons from now. The value only appears with age — which is exactly why you have to build it before you need it.

Try it

3.1 — Find a hub in your own life. Think of a person you'd mention in a lot of notes. Partner, boss, kid, closest friend.

Now count roughly how many unrelated areas of your life they'd appear in. Health. Money. Travel. Work. Weekends.

That's a hub. In a linked-notes memory, that one node connects all of those areas to each other. Ask about your dentist and you can reach your mortgage in two steps. That's the decay, and you just found it in your own life without writing any code.

3.2 — Predict the fix. Before the next lesson: given that hubs are the problem, what would you do about them? Delete them? Cap their connections? Something else?

Write it down. Course 09 has the real answer, and there's a good chance you'll get close.

What's next

Everyone lists four jobs a memory has to do. There are seven, and the three that get left out are the entire difference.