The best practice that builds the wrong memory

Ask an engineer to build a memory for a machine and watch what they reach for first. Tidy categories. A type for every record. A clean place for each thing to live. It’s good instinct, drilled in over years, and it has shipped a thousand working systems.

It’s also why almost every AI memory falls short the same way.

The thing every memory system has in common

Look at how these systems store what they know. One files each item under a label. Another sorts entries into types. A third tags everything so it can be found later. The surface differs. The move underneath is identical: decide what each piece is, then put it in the right drawer.

That move feels like rigor. Filing is what a careful engineer does. And for most software it’s exactly right, because most software handles things that genuinely have types.

A memory of how you think isn’t one of those things. A single thought is about four subjects at once, connects to a dozen others, and means something different next week than it did today. Put it in a drawer and you’ve cut every thread that ran out of it. The drawer was the mistake, not the filing of it.

Where the instinct comes from

Here’s the part worth sitting with, because it explains why this keeps happening.

A system trained on how software is normally built has seen the clean version a million times. Status fields. Typed records. Service boundaries. Normalized tables. So when you ask it to help build a memory, the path of least resistance runs straight back to the patterns it has seen most. It doesn’t propose a messy web. It proposes a neat hierarchy, confidently, because a neat hierarchy is what “good engineering” has always looked like.

I know this firsthand, because I spend a real part of my week stopping my own coding assistant from tidying up. It keeps offering me one more category, one more clean type to sort things into. Every time, it’s sure it’s helping. Every time, it’s quietly rebuilding the exact filing cabinet the whole thing exists to avoid. I’m not debugging the code. I’m arguing with a reflex.

The tell is always the same shape: a clean hierarchy appears where there should be a tangle. When the structure looks too tidy for the mess it’s supposed to hold, the old instinct has crept back in.

So copy biology? Not blindly

The obvious answer is to stop copying engineering and start copying the brain. The brain doesn’t file memories in drawers. It holds them in a web where everything can reach everything, and it works far better than anything we’ve built.

But “just copy the brain” walks into its own trap. The brain is full of limits that aren’t clever design at all. They’re workarounds for running on cells. It can only hold a few things in mind at once. It can’t cheaply mark something as deliberately absent. It rewrites its own past and keeps no clean record of the change. Copy those faithfully and you’ve imported a set of problems biology never chose and would drop in a heartbeat if it could.

So neither instinct wins on its own, not the engineer’s drawers and not biology’s limits copied whole.

The line that actually sorts it

There’s a cleaner question than “brain or machine,” and it does the real work.

When you find a feature of the brain, ask what it is. Sometimes it’s function: the brain works this way because the problem genuinely calls for it. Holding memories as an open web is function. That’s worth copying, every time.

Sometimes it’s a workaround: the brain works this way because cells are slow, space is scarce, and energy costs something. The few-things-at-once limit is a workaround. So is the inability to keep a perfect record. Those aren’t worth copying, and here’s the gift: fixing them is nearly free. The whole reason the brain settled for the workaround was a constraint software doesn’t have. Software can keep everything, mark true absence, hold a perfect history. The bug was never in the function. It was in the wetware, and we left the wetware behind.

Copy the function. Fix the workaround. That single line tells you when to follow biology and when to overtake it, and it cuts straight through the false fight between the engineer’s drawers and the brain’s limits.

What this asks of you

Best practice earned its name on problems that have types. A memory of your own thinking is not one of those problems, and the practices built for filing cabinets will quietly rebuild a filing cabinet inside it, no matter what you meant to build.

The way past it isn’t a better category. It’s the willingness to leave the mess messy where the mess is the point, and to fix biology’s old limits where they were never the point at all. The instinct that tidies is the one to watch. Most days, it’s the smartest-sounding voice in the room, and most days, for this one problem, it’s wrong.