The Human Is Doing the Remembering

Years ago, in a government office, a colleague and I made a small bet. His secretary was filling out a purchase order for a box of business cards — something well under the dollar threshold that was supposed to require one. I said I knew why she was doing it. He didn’t believe me. We walked into her office and asked. She didn’t hesitate: “Because that’s the way we’ve always done it.”

He owed me dinner. I didn’t think about that exchange again for thirty years.

Then, a few weeks ago, in the middle of a conversation about something completely unrelated — why a set of AI interaction primers I’d built had quietly split into two documents without anyone questioning whether that split still made sense — the secretary showed up again. Uninvited. Unsearched for. She just arrived, fully formed, at the exact moment she became useful.

I didn’t go looking for her. I wasn’t thinking about government procurement, purchase orders, or 1990s office politics. I was thinking about primer architecture. The only thing connecting the two situations was the shape underneath them: a process that had outlived its original justification, running on inertia because nobody had asked why it was still running.

That arrival — instant, unprompted, exactly on time — is the seed of a question I’ve been circling for weeks. It’s about something I think most people who use AI regularly have started to feel, even if they haven’t had a reason to name it yet.

What happens when the work of maintaining continuity in a conversation quietly shifts from the system to the person using it?

The Assumption Nobody States

Every conversation — human or otherwise — has to maintain some kind of coherence over time. Someone has to remember what was said, notice when things have drifted, and reconnect the present moment to what came before. In conversations between people, that work is roughly shared. Both sides are doing some of it, mostly without noticing.

In conversations with AI systems, that work isn’t shared. It’s yours.

The AI responds to what’s in front of it. It doesn’t carry a evolving sense of who you are from one session to the next unless you bring that with you. Which means you are the one who has to:

Remember what’s already been established, so you don’t re-explain it.

Notice when the conversation has drifted from what you actually meant.

Reintroduce context, vocabulary, and prior observations at the right moment.

Correct misunderstandings before they compound.

Decide when a conversation has run its course and what needs to carry forward into the next one.

None of this feels like a special task. It feels like using AI. People who are good at it are described as having strong “prompting skills.” But what they’ve actually developed is something closer to fluency in carrying their own continuity — quietly, constantly, without being asked, and without it ever showing up as a line item anywhere.

I started calling this continuity labor. And once I had a name for it, I started seeing it everywhere — including in my own work. The primers I’d built, the carryover documents I write at the end of long sessions, the “here’s where we left off” summaries — all of it was continuity labor, made visible because I’d built tools to manage it. Most people are doing a version of the same thing without ever building the tools, which means without ever seeing the labor as labor at all.

Storage Is Not the Problem

The dominant response to “AI doesn’t remember enough” has been to build bigger memory — larger context windows, vector databases, retrieval systems that pull in relevant documents before responding. These are real improvements. But I think they’re aimed at the wrong target.

There’s a difference between retrieval and relevance determination, and most current systems are optimized for the former while the actual bottleneck is the latter.

Retrieval answers: what information is available that relates to this topic?

Relevance determination answers: which prior thing — out of everything that’s ever happened — actually matters right now?

These are not the same question, and the difference is the whole paper.

Go back to the secretary. If you handed a modern AI system a transcript of my conversation about primer architecture and asked it to search for relevant prior material, it would almost certainly not surface a thirty-year-old story about purchase orders and business cards. There’s no keyword overlap. No topical similarity. A purchasing anecdote from a government office has nothing to do, semantically, with AI primer design.

And yet it was exactly the right thing to surface. Not because of what it was about, but because of what it was — a process running on inherited justification, unexamined, persisting because nobody asked why. That’s structural isomorphism: two things that look completely unrelated on the surface, connected by an identical underlying mechanism.

My own memory found that connection instantly, with no effort, no search query, nothing I’d recognize as “thinking.” A relevant example simply arrived, already evaluated, ready to use. A system built on semantic retrieval would need the content to match. What actually mattered was that the mechanism matched — and mechanism-matching isn’t something retrieval systems are built to do.

What This Explains

Once I had this distinction — retrieval versus relevance, storage versus structural matching — a lot of things I’d been treating as separate observations turned out to be the same observation wearing different clothes.

Prompt engineering, as a skill, is mostly continuity labor in disguise. The “right” prompt is so often just the prompt that supplies, upfront, the context the system can’t retrieve or infer on its own. People who are good at this have gotten good at predicting what the system will need and pre-loading it — which is a real skill, but it’s a compensation, not a capability the system has.

Primers — documents written specifically to restore context at the start of a session — are an even more explicit version of the same compensation. The fact that primer-writing has become a recognizable practice, independently developed by different people across different platforms, tells you the underlying need is structural. Nobody invents the same workaround independently unless the problem they’re working around is real and shared.

Drift correction — the constant, low-grade work of noticing when a conversation has started responding to a generic version of “people in general” rather than to you specifically, and pulling it back — is continuity labor too. It happens so often in long conversations that experienced users barely register it as a distinct action anymore. It’s just part of “talking to AI.”

None of these are hacks or quirks. They’re adaptations. And adaptations are evidence. When a lot of people, independently, develop similar workarounds for the same friction, that friction is telling you something about the structure of the system — not about the people.

The Externalization

Here’s the sharper version of the claim, the one that I think matters most: current AI systems don’t just lack the ability to carry continuity. They externalize the cost of continuity onto the human.

The work doesn’t disappear because the system can’t do it. It moves. It moves onto the person typing. And because it moves quietly — distributed across small, individually unremarkable actions like rephrasing a question or adding “as I mentioned earlier” — it never gets counted. It just becomes “how you use AI.”

This matters because it changes the question. The usual framing is: how do we make AI remember more? But if the real cost is continuity labor — the ongoing work of figuring out what’s relevant right now, given everything that’s happened — then more storage doesn’t necessarily help. A system that can hold ten million tokens but still requires you to figure out which of those ten million tokens matter hasn’t reduced your workload. It’s just made the haystack bigger.

The better question might be: how do we reduce the amount of this work a person has to do, without taking away their ability to direct the conversation themselves?

That second half matters. The goal isn’t an AI that decides what’s important for you. It’s a system that does more of the noticing — surfaces a structurally relevant example the way my own memory surfaced the secretary, unprompted — while leaving you free to use it, ignore it, or correct it.

A Test in Real Time

What’s a little strange about this paper is that the process of writing it became an example of its own subject. Across the conversations that led to it, I worked with multiple AI systems, each carrying different pieces of context, each occasionally drifting toward generic explanations before being pulled back to the specific thing I was actually pointing at. At one point, in the middle of evaluating a cover image for this very paper, one of those systems started discussing the image as if it were an academic publication cover — which, structurally, it was — and began offering Zenodo-related observations I’d never asked for. Nothing was remembered incorrectly. The wrong association simply became active at the wrong time, exactly the kind of thing the paper describes.

I didn’t plan that. It just kept happening, the way Sharon kept showing up — relevant, on its own schedule, demonstrating the point by being the point.

Where This Leaves Things

This paper doesn’t claim AI systems are failing at their jobs. It claims something more specific: that the way we currently talk about “AI memory” focuses on the wrong layer of the problem, and that a significant, mostly invisible amount of work has been quietly handed to the people using these systems — work that looks, from the outside, like nothing at all.

If you’ve ever found yourself re-explaining something you’re sure you already said, or rephrasing a question for the third time because the conversation keeps sliding sideways, or building yourself some version of a “primer” just to get a fresh session oriented — you’ve been doing this work. Probably for a long time. Probably without a name for it.

It has a name now. Whether that turns out to matter is an open question. But Sharon, the secretary, the purchase order, and the business cards suggest that some questions are worth asking even thirty years late.

Continuity Burden Asymmetry: The Externalization of Continuity Labor in Human-AI Interaction DOI: 10.5281/zenodo.20673479

Companion to Reflective Parallax and the Emergence of the Transformed Interpreter DOI: 10.5281/zenodo.20580335

Michael Todd Magee