The Problem Nobody Tells You About

Here is the sentence the rest of this course exists for:

A memory that saves everything and deletes nothing gets worse over time.

Not slower. Worse. Less able to answer the question you asked.

That sounds wrong, so let's do it with something you already own.

1. Your phone

Search your phone for that text about your car.

Back when you had 200 messages, you found it in a second. Now you have 40,000, and the same search hands you a wall of results — your mechanic, three group chats, a car ad someone forwarded, your sister asking for a lift — and the one you actually wanted is somewhere on page six.

Your phone isn't broken. Search did exactly what it always did. It looked for messages about cars and gave you messages about cars.

You just have more stuff, so "close enough" got a lot worse.

Nobody sent you a warning the day it tipped over. There wasn't a day. It got slightly worse every month for four years.

2. Now the same thing, with an AI memory

Say you save a note:

Insulin goes in the fridge, 36–46°F. Never freeze it. Throw it out 28 days after opening.

Good note. Worth keeping.

At 50 saved notes — you ask "how do I store insulin" and you get it. First result. Done.
At 500 — about thirty notes now mention insulin somewhere. Old shopping lists. A calendar reminder. A note that just says "pick up insulin." The real answer is still in there, but it's competing.
At 5,000 — the real answer isn't in the top five results at all. The AI reads whatever it was handed, which is now five notes that mention insulin and don't answer the question, and it gives you a confident, useless reply.

 how good
your answers │████████████▄▄▄▄▄▄▄▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁
      are    │              ╲
             │               ╲___ every note you add makes every question worse
             └──────────────────────────────────────►  how much you've saved

Nothing broke. Nothing threw an error. Nothing appeared in a log.

The system asked for the closest matches and got the closest matches. There were just more things competing to be closest.

3. The part that should worry you

There is a point where your AI is worse with a memory than without one.

Past that point, the memory is actively damaging the thing it was built to help. An AI with no memory at least says "I don't know." An AI with a bad memory hands you the wrong answer with total confidence, because from the inside, five mediocre matches look exactly like five good ones.

And almost nobody notices the moment they cross it. There's no alarm. People just feel it slowly go from "wow, this is amazing" to "this used to be better," and they can't say when it changed or why.

Then they blame the model, and go pay for a bigger one.

4. Why a bigger computer doesn't help

This is the bit worth being clear about, because it's where money usually goes.

A faster machine searches the same 5,000 notes faster. It returns the same wrong five. Speed was never the problem.

A better search engine picks slightly better matches from the same 5,000. It helps a little. It does not fix the shape of the problem, because the problem is that too many things are legitimately relevant and only a handful fit in the stack of paper.

A bigger context window — a taller stack — lets you hand over more notes. Now the AI reads twenty mediocre notes instead of five. That is not obviously better, and it costs more every single time.

This is a growth problem, and growth problems are not fixed by going faster.

5. So here's the reframe

The problem isn't forgetting. It's keeping everything.

Your own head does not store every moment you have ever lived.

You have forgotten what you had for lunch three Tuesdays ago. You have forgotten almost every journey you've ever taken to work. That isn't a failure — it is the reason you can find anything in there at all.

Your brain drops what it doesn't use, constantly and aggressively, so the things that matter stay reachable. A brain that kept everything at equal strength would be a brain that could not answer a question.

So the real question this course answers is not:

~~How do I give my AI a memory?~~

It is:

How do I teach it what to let go?

That's a much harder question. It's also the one nobody else is answering, and it's why what you build here will still work in year two.

Try it

Nothing to install — you're going to look at your own life.

2.1 — Find your own tipping point. Open your photo library. Try to find a specific photo from about three years ago. Not the newest, not a favourite — a specific one you remember taking.

Notice how you actually do it. You don't search. You narrow: roughly what year, roughly what season, what else was happening then. You use connections, because the pile is too big for matching alone.

Hold onto that. Course 04 builds exactly that ability, and it's the reason plain search isn't enough.

2.2 — Write down a guess. Before the next lesson, answer this in one sentence:

If your memory has to throw things away to stay useful — what should it throw away first?

Keep your answer. Course 09 is entirely about that question, and the honest answers are not the obvious ones.

What's next

Three ways people actually build these things, and what each one costs you. Two of the three are what you'll find in every tutorial. Both of them decay.