What You Choose to Remember Shapes What You Can Do Next

AI can carry the memory. It should not decide what matters.

Anyone who has used AI for a while knows the pattern. The first stretch is strong. You set up the project, hand over the context, and the work comes back sharp and useful. Then, gradually, it dulls. The answers drift towards the generic, and the output you are getting now is a little flatter than what you were getting a month ago. It seems like it had a few too many drinks last night.

So we reach for the obvious answer. We give it more to hold, what we believe to be "more context". We dump bigger files at it, longer documents, whole artefacts, folders of past work, more standing instructions, more notes. The assumption underneath is that volume is the fix. If the work is going flat, the tool just needs more in front of it.

It's the wrong fix. The problem was never how much the tool remembers. It is what ends up in that memory. What’s missing is the intentionality of deciding what matters in that memory. 

Context is not volume. It is relevance, chosen with intent.

Think about how a person gets better at anything. Not by remembering everything, because no one can. A brain holds only a fraction of what passes through it, and has to be intentional about what it keeps. That selection is not a limitation we work around. It is the mechanism. The reasoning that stuck, the lesson we chose to carry, the pattern we noticed and held on to, is what lets each round of work start a little further along than the last. 

We have compounded knowledge this way for as long as we have done work at all, throughout the history of mankind. This is how civilisation has been built over time. The intentional curation and documentation of compounding knowledge. Logic, reasoning, lessons learnt. We get better by choosing well what to keep, and discarding the rest. 

It is tempting to think an AI's memory runs on different rules to ours. That because it can hold so much more, it no longer needs to be selective, and the discipline of choosing can be retired. The opposite is true. The more that can be kept, the more that which is worthless if also kept, and the more it matters who decides what belongs. Storage went up. The principle did not change. If anything it matters more now, not less, and we have responded by handing the choosing to the tool. We pour volume in, rarely tell it what worked or why, and trust it to choose well.

We are abdicating the choosing: the judgement of what mattered, what was worth carrying forward, what the next piece of work should be built on. That judgement was always the human part, and it is the part we should be least willing to give up.

It is easy to read this as a discipline lost to laziness, a good habit we let slip once the tool made it optional. But that would not be true. The deliberate version of the discipline was barely there to begin with for most people. Most people have never written up the reasoning behind a decision, the option they rejected, the thing a piece of work taught them. Some people do this loosely but naturally in journaling, through intentional reading and annotation, and other such ways. Some were intentional enough to structure this down into published writing, and we have the get to read it today to shape our thinking and build of their work. 

But overall, we got away with it because the doing was its own kind of keeping. Working a problem through, drafting it, reasoning it out, was how the lesson lodged in the first place. We retained by doing, not by recording. The recording we mostly skipped. The doing quietly did the work of making us sharper next time. Doing the work is how our minds filtered what was important to keep, and what should be discarded.

That is the part now changing. We are starting to hand the doing itself to AI. The drafting, the synthesis, the working-through. And when the doing goes, so does the keeping that came free with it, the retention we never realised we were getting. The deliberate discipline was never built to fall back on. So we lose both at once. The recording we never did, and the keeping we did not know we relied on.

This is, I think, much of what sits underneath the worry about deskilling. It tends to get framed as losing the ability to do the task. The deeper loss is quieter. 

Doing the work was how we compounded, and the choosing of what to keep was how we steered that compounding. Hand off the doing, and you stop getting better at the work on your own. Hand off the choosing as well, and there is no longer a human at the centre of any of it. If the tool does the work, and also decides what the work was worth, it is fair to ask what we are doing with our time.

There is a further problem, and it is the one most people never see. Once the choosing is handed over, a memory assembles itself in the background, holding whatever the model judged worth holding. We trust it because it was the one that kept it, and we stop treating it as something to check and start treating it as something true.

It looks like knowledge. It behaves like a graveyard.

We have done the harder version of this before, just not for knowledge. For more than a decade, organisations have built data architecture, and underneath the tooling, data architecture is a discipline of deciding. What goes in. What is clean. What is authoritative. What can be trusted to reason from. Nobody serious points AI at a raw, ungoverned data lake and trusts the answer. And yet that is close to what we are doing with everything else that we know. 

We point the tool at all of it, let it keep what it likes, and hold thumbs that it chose well. The question we are mostly avoiding is the obvious one. What if it didn't?

Valentin Marguet wrote a clear, practical post recently about running a project workspace in four folders. Context, work, library, and knowledge. The first three keep you organised, what you do, how you do it. The fourth, the one holding the lessons learnt, the retro, the small rule that says if this happens, do that, is the only one that slowly makes you better, and the one most people skip. And importantly, the intentionality of deciding what goes into that fourth folder. 

The gap he spotted in individual work is the same gap that runs through whole organisations. A great deal of effort goes into recording what we know and what we do. Well, that might be an over assumption for many organisations. Almost none goes into how we think, what we have learnt, and the choices we made and why. Which is exactly why so much of it ends up in a few senior heads, and leaves when they do.

Now coming back to our new tooling, the instinct, once the problem is visible, is to go back and tidy the memory. Review what the AI kept, correct it, prune it. Better than nothing, but it comes too late. 

But three months on, you can no longer reconstruct what was worth keeping in the moment, so you end up checking the tool's choices against the tool's choices. 

The judgement has to happen close to the work, while you still know what mattered and why. That makes it a decision at the point of creation, or near it. “This is worth keeping”. “This is what it means”. “This is where it goes”. Done that way, your record becomes the authoritative one, and the tool's memory becomes the thing you check against it, asking what is missing rather than trusting what is there.

None of this is an argument against the tools. It is almost the opposite. If AI is going to compound human capability rather than quietly replace pieces of it, we have to become much more precise about what we hand over.

The tools can carry more than we can. It can capture, structure, retrieve, compare, summarise, and draft at a scale no person should try to match. That is the gift in it.

What it should not be handed, is the deciding; what is worth keeping, what is true, what is authoritative, what your future self and your team and your agents will reason from. What should become the basis for the next decision, the next piece of work, the next act of judgement, that is the human part.

Not the decorative human part. The pertinent one. The part where experience becomes understanding, and understanding becomes something we can build on. If we hand that over silently, we should not be surprised when the work keeps moving, but neither the tool nor the human behind it gets meaningfully sharper.

Let AI carry the memory. Do not abdicate the meaning.