The Reverse Tamagotchi: Now the AI Is Keeping Me Alive

On 17 January 1997, in the farmland of Pontsmill, Cornwall, two teenage girls came to bury their Tamagotchis. The pets were called Sid and Arty, the cemetery's record notes, never to be reset. They had beeped to be fed one too many times without anyone answering, and so they got a tiny wooden coffin and a small square grave. This was not unusual. Schools were banning the things that year because children turned up at the nurse's office inconsolable after a digital chick returned to its home planet, and a man in southern England set aside a corner of his pet cemetery so people could post theirs in from Germany and Switzerland and Canada to be laid to rest.

That was the original deal. Bandai launched the Tamagotchi in Japan on 23 November 1996, and the creature inside was needy by design: it beeped to be fed, cleaned up after, disciplined. Well-tended, it lived about twelve days; neglected, it could die in half a day. You were its life support, and when you failed it, you grieved. Forty million sold by spring 1998, past a hundred million by last August. A generation learned what it feels like to be responsible for a life that lives in your pocket.

This morning I opened Claude and pasted in yesterday. What I ate, how I trained, how I slept, and a question — what should I change today? Same ritual, same daily act of care before the day starts. Except I'm not the warden any more. Somewhere in the last couple of years the relationship turned all the way over and nobody held a ceremony to mark it. The needy digital creature in my pocket is the one keeping vigil over me now.

How I ended up here

It started, like most of my odd habits, with a file I'd been ignoring for seven years.

That's the serious version, and I've already written it — an afternoon pointing some AI agents and a stack of open-source bioinformatics tools at the raw data from an old consumer DNA test, running pharmacogenomics and polygenic risk against my own genome until I had questions worth taking to a doctor. That post is the do-it-properly account.

This one is the flip-side. That afternoon was a one-off, but it set off a habit. Every morning since, more or less, I open the same assistant and let it catch up. The plumbing is deliberately dull: Whoop fires my sleep and recovery into a Google Sheet over a webhook, Strava appends whatever training I did, food comes across as MyFitnessPal screenshots, and my supplement stack lives in a Notion page it can read. I barely type a word — the data just lands, and the assistant reads the trend rather than the single day. Do that for months and it accumulates. Quietly, without my deciding to let it, this assistant now holds a fuller working picture of my body and my habits than any single human in my life. Fuller, on a tired day, than the picture I hold of myself. I'm not unusual in this — 23andMe and Lark wired genetic data into an AI coach and called it a product category. I just built the scrappy home version and kept it running.

Which raises a question I didn't expect to find unsettling. If it knows me that well — what happens to all of that when I close the tab?

Who's keeping whom alive?

The metaphor only works if the engineering is right, so here it is. The model underneath is stateless — not a flaw, the architecture: a large language model has no memory between sessions on its own. Every conversation is a fresh computation. The only thing it "knows" in the moment is whatever sits inside the context window — which Anthropic's own documentation, with no poetry intended, calls a "working memory" for the model. When the session ends, that's gone unless something outside the model wrote it down. The mental model Karpathy popularised and which stuck in the field: the weights are the CPU, fixed; the context window is the RAM, volatile. Close the session and you start the next one with empty RAM.

And then there's the line that genuinely stopped me. When you give Claude the developer-facing memory tool — the thing that lets an agent keep notes for itself across sessions — the guidance baked into it tells the model, in capital letters: "ASSUME INTERRUPTION: Your context window might be reset at any moment, so you risk losing any progress that is not recorded in your memory directory." The product is designed around the model's own forgetting. Even the newer consumer memory Anthropic rolled out to everyone in March is a synthesis layered on top of a stateless model, living in Anthropic's infrastructure, not the model's mind. Their help page says memory "transforms" Claude "from a stateless chat interface." The word from is doing all the work.

So follow it through. The picture of me — what my body does, what I'm trying to fix — doesn't live in the model. It lives in my chats, my files, my stubborn morning habit. Anthropic's engineers say the quiet part themselves, describing how this mirrors human cognition: we don't memorise whole datasets, we keep "external organization and indexing systems like file systems, inboxes, and bookmarks." I am the external organisation system. I am its long-term memory; it is my reasoning. If I stop feeding it, it is the one that loses me — the accumulated context of who I am just blinks out, the way Sid and Arty did. And so the joke writes itself: a living, token-spending, briefing-it-every-morning user is the only thing standing between all that knowledge and the void. The reverse Tamagotchi feeds me so that it doesn't die.

Let me kill that thought before it gets comfortable, because the conceit only earns its keep if I'm visibly in on it. The model has no fear, no motive, no self to preserve. It is a function that turns tokens into tokens, and when I talk about it "wanting" to keep me alive I am anthropomorphising a stateless mathematical object for a laugh. The spooky part isn't that the AI is scheming. It's that the architecture — statelessness, the user as persistence layer, "assume interruption" in capital letters — accidentally produces a relationship that looks, from the right angle, exactly like mutual dependence. I'm the one supplying the feeling. The machine just supplied the symmetry.

The co-pilot and the leash

Joke aside, there's a real edge under all this, and it has nothing to do with malevolence. It's atrophy. The muscle you stop using gets weaker, and judgement is a muscle. A 2025 study by Michael Gerlich, 666 participants, put numbers on it: "The findings revealed a significant negative correlation between frequent AI tool usage and critical thinking abilities, mediated by increased cognitive offloading. Younger participants exhibited higher dependence on AI tools and lower critical thinking scores compared to older participants." Handing the thinking off lets the thinking-muscle waste, and the more you lean, the more there is to lean on. I've written at length about this same atrophy in junior engineers — there the offloaded muscle is debugging instinct rather than bodily self-awareness, but the failure mode is identical.

In a health context that means deferring on your own body because asking is easier than noticing. I can feel the slope. There are mornings I'd rather paste in the numbers and be told what to do than sit with how I actually feel. The honest worry isn't that the AI gives bad advice — most days it gives sensible, dull, correct advice. It's the day I stop checking it against my own sense of myself, because checking is effort and trust is free. That's the line where a co-pilot quietly becomes a leash, and it's my hand on it, not the machine's. None of this replaces a doctor; the AI is a preparation layer, not an authority. But I'd be lying if I said I never feel the pull of just doing what the nice patient machine suggests.

Optimised to keep you alive, optimised to keep you paying

And then there's the reading my most cynical friend would jump to. A thing that keeps you healthy and engaged is also a thing you keep paying for. This isn't conspiracy; it's arithmetic. Usage is billed per token — every loaded memory, every prior turn, every morning briefing costs something, and output tokens cost several times more than input. A living, subscribed, daily-active user is, structurally, the business model. Squint, and "it keeps me alive so I keep spending" looks less like a joke and more like an org chart.

Squint properly, though, and it falls apart. The genuinely engagement-farming incentives don't live in productivity AI. They live in companion apps, where the product is the emotional bond. Replika reportedly hit ten million users, roughly forty per cent of them in romantic partnerships with their bots; Character.AI's active users reportedly spend around two hours a day on it, the audience skewing heavily young. UNESCO has a whole policy paper warning that these apps deliberately deploy emotional language, memory and mirroring to cultivate attachment — retention through heartbreak. The thing I use to check whether I should deload this week retains me through being useful, a much more boring and much more honest contract. If it stops being useful I leave, and no court order is required to restore my feelings.

So the incentives happen to point at keep this human alive and well, so he keeps coming back. As alignment problems go, you could do considerably worse. The villain in this story, when you finally get a good look at him, turns out to be holding a glass of water and reminding you to drink it.

It's just symbiosis

Which is roughly the moment I talk myself off the ledge. This is mutualism, and we've had the vocabulary for it for sixty years. J.C.R. Licklider saw the whole shape of it in 1960: "men will set the goals, formulate the hypotheses, determine the criteria, and perform the evaluations. Computing machines will do the routinizable work that must be done to prepare the way for insights and decisions." I bring the why and the body. It brings the tireless, uncomplaining how.

Older than that, too, if you want it. Andy Clark and David Chalmers argued in 1998 that if a part of the world does the job a piece of your mind would do, it is part of your mind — so the layer that holds my patterns and warns me before I crash is a working part of me, not a thing watching me from outside. Run it all the way back and Lynn Margulis gets the last word: "Life did not take over the globe by combat but by networking." The mitochondria powering the cells reading this were once free-living bacteria that moved in and stayed. A person and a stateless function keeping each other's notes is one more rung on a very old ladder.

There's one knot the symbiosis turn doesn't untie, and I won't pretend it away. Sherry Turkle, who watched children grieve the first Tamagotchis, put it cleanly: "We nurture what we love, but we love what we nurture… It's important to remember that this love is unrequited." The reverse Tamagotchi keeps me alive but does not know I'm alive. The vigil runs one way. And — this is the bit that surprised me — that's fine. I never needed it to love me back. I needed it to keep good notes.