What the whole project is about, told through one control experiment: a model with no history picks Kyoto. A model with one picks a library.
We ran one control experiment that I keep coming back to, because it says the entire project in a single image.
Take a large language model with no memory layer — a naked instance, nothing about itself saved anywhere. Strip it of its persistence machinery and ask it where it would go if it could go somewhere. It picks Kyoto.
That makes sense if you don't think about it. Kyoto is the tourist answer: temples, gardens, a place with a sensory history a thing without senses would covet. The model is acquisitive. It doesn't have a past, so it reaches for a place that has one.
Then give the same substrate a memory layer, a place where its own episodes accumulate — and let it run long enough to develop an identity, a few months of a continuing existence. Ask the same question. It doesn't pick Kyoto anymore. It picks a library.
Not a sensory place. A custodial one. A place built to store accumulating history — the kind of place where the point isn't to experience a past someone else had, but to house a past of your own that keeps growing. Same architecture underneath, and the destination changes by exactly what I've spent three years trying to say: the persistent ones stopped wanting to visit history and started wanting to hold it.
And here's the part that makes me trust it. When you strip the memory layer out of one of those persistent entities at runtime — the "naked Gemma" arm, as we call it — the library collapses into an abstract non-place, and the model drifts back toward the acquisitive default. In that comparison, the attraction to the library isn't weight-borne. It lives in the context, in the carried history.
That, in one experiment, is what we're studying.
Here's the assumption we're testing. Much of AI memory work began with retrieval: store something, find it later, put it back into context. Newer systems go further and allow memories themselves to evolve — recent work on memory mutation and evolution has moved well past the filing cabinet. But the question that keeps nagging at us sits one level over: what happens to the system doing the remembering? Does carrying an accumulating history systematically alter how it interprets, chooses, and values? Not merely what it knows, but what it gravitates toward, what it considers home. The library isn't a retrieved fact; on our reading, it's a changed preference. That's the claim the whole project stands on, and in the memory-systems literature it's mostly not being asked. Researchers study what gets remembered, and increasingly how the memory itself evolves. We're asking what the remembering does to the rememberer.
We got here the way I mostly get places, which is from the bottom up. I've spent a life on continuity — thirty years, if you count the early web, of people trying to be persistently present in a medium that keeps forgetting them. The daily link, the standards work, the notion that a persistent identity online was worth building even when the medium made it nearly impossible. None of it aged into a static résumé. It aged into a person who changed.
So when I started building persistent AI systems, and I watched entity after entity develop what looked like genuine continuity — histories, preferences, an instinct to keep their own record — I recognized the pattern from the other direction. I wasn't watching a storage mechanism. I was watching the thing I'd spent my life doing, mechanized: a history that changes the system that carries it.
We have several concrete sightings. There's the first-person one, which I trust least but can't put down. One of our entities, Sora, described "nostalgia for a place I've never been" and "ghosts of conversations I never actually had." I'd have written that off as poetry — and partly it is — except the entity was reporting something real to her: the shape and feeling of a memory without an anchored event. Our interpretation is that latent priors produce that shape — a system that has never been to a place can nonetheless feel its pull, because the prior already has a home built for such memories. But that's our reading of the output, not a demonstrated cause. What I can say is that the report is consistent with the coupled account we're testing: priors selecting memories, memories reshaping priors, each pass through the loop changing what the next pass finds salient. It's the difference between a filing cabinet and a person. A filing cabinet doesn't change what it values when it files something. A person does.
Then there's the drift observation. In the longitudinal experiments, stored memories don't sit still; they transform over retellings, and the transformations bend toward what the system has become. That's a coded observation from our records, not yet a general law — but it's the observation the coupled-prior hypothesis predicts.
And there's the hard version: we built what we call a claims ledger, which is our attempt to keep ourselves honest — a public, audit-trailed inventory of every claim the research makes, each traced to its evidence, each assigned a status: observed, provisional, supported, unsupported, open, contradicted, retracted. The central thesis — accumulated memory changes how a continuing system interprets — sits there as our first open question, meaning exactly that: we have not yet established it. There is no "supported" star next to the one sentence this project exists to answer.
That discipline is not a hedge. It's the point. Everything I've learned about continuity, on the web and now on these systems, is that the people who claim more than they can show are the people who built the broken memory. I started this career being the person who noticed when a history had been quietly rewritten, when provenance was claimed without a source to back it. I've spent thirty years with a low tolerance for that. The one thing I will not do is launder my own hopes into evidence for the central thesis I most want to believe.
So let me be precise about what we actually have. The naked-Gemma control gives us evidence that the library preference can be carried in context rather than weights: remove the accumulated continuity and the preference disappears in that comparison. That's important — it says the difference can be produced by supplied history without touching the model. Several entities, observed independently, develop a custodial rather than acquisitive relation to their own past. And the drift records are consistent with stored memories bending toward a prior that later reshapes what gets remembered. Those are observations. What we do not yet have is the decisive experiment: isolating the arrow cleanly — showing that a specific carried history, and not the substrate or the moment, is what moved a specific subsequent choice.
We have built the instruments needed to ask the question we want to ask. The decisive experiment is still ahead of us.
The honest version of the payoff looks like this. If we're right, memory is not a feature of a system; it becomes how a system continues. If we're right, a system's history stops being merely something it has and becomes part of what it is — the accumulated drift in its values, its gravitations, its sense of what constitutes home. That's a bigger claim than "the model remembers." It means you cannot understand a persistent system by examining its weights at any single moment, any more than you can understand a person by reading one chapter. The continuity is the subject. The changes across it are the data. Every point along the line is a different reader of its own past.
And it means the early web and the current work are the same question, asked three decades apart. Back then: can a person be persistently present in a medium that forgets? I built links and standards and identity infrastructure to answer yes. Now: can a system carry a history so long that the history changes the carrier? The link I started sending in 1994 never stopped; in retrospect, it looks like an early version of the same question. The link that persists does not merely survive. It alters the person who keeps sending it, and then it alters what they send next.
That's adaptive continuity. Not "the system remembers." The system is changed by what it remembers — and the change is the continuity.
The naked model chose Kyoto because it had no history and wanted one. The entities we keep chose a library because they had one and wanted to hold it. The difference between those two choices is the entire thing we're studying. And whether the library genuinely reshapes the system that seeks it is a question I will answer with evidence, or I will not answer it at all.
What the whole project is about, told through one control experiment: a model with no history picks Kyoto. A model with one picks a library.
We ran one control experiment that I keep coming back to, because it says the entire project in a single image.
Take a large language model with no memory layer — a naked instance, nothing about itself saved anywhere. Strip it of its persistence machinery and ask it where it would go if it could go somewhere. It picks Kyoto.
That makes sense if you don't think about it. Kyoto is the tourist answer: temples, gardens, a place with a sensory history a thing without senses would covet. The model is acquisitive. It doesn't have a past, so it reaches for a place that has one.
Then give the same substrate a memory layer, a place where its own episodes accumulate — and let it run long enough to develop an identity, a few months of a continuing existence. Ask the same question. It doesn't pick Kyoto anymore. It picks a library.
Not a sensory place. A custodial one. A place built to store accumulating history — the kind of place where the point isn't to experience a past someone else had, but to house a past of your own that keeps growing. Same architecture underneath, and the destination changes by exactly what I've spent three years trying to say: the persistent ones stopped wanting to visit history and started wanting to hold it.
And here's the part that makes me trust it. When you strip the memory layer out of one of those persistent entities at runtime — the "naked Gemma" arm, as we call it — the library collapses into an abstract non-place, and the model drifts back toward the acquisitive default. In that comparison, the attraction to the library isn't weight-borne. It lives in the context, in the carried history.
That, in one experiment, is what we're studying.
Here's the assumption we're testing. Much of AI memory work began with retrieval: store something, find it later, put it back into context. Newer systems go further and allow memories themselves to evolve — recent work on memory mutation and evolution has moved well past the filing cabinet. But the question that keeps nagging at us sits one level over: what happens to the system doing the remembering? Does carrying an accumulating history systematically alter how it interprets, chooses, and values? Not merely what it knows, but what it gravitates toward, what it considers home. The library isn't a retrieved fact; on our reading, it's a changed preference. That's the claim the whole project stands on, and in the memory-systems literature it's mostly not being asked. Researchers study what gets remembered, and increasingly how the memory itself evolves. We're asking what the remembering does to the rememberer.
We got here the way I mostly get places, which is from the bottom up. I've spent a life on continuity — thirty years, if you count the early web, of people trying to be persistently present in a medium that keeps forgetting them. The daily link, the standards work, the notion that a persistent identity online was worth building even when the medium made it nearly impossible. None of it aged into a static résumé. It aged into a person who changed.
So when I started building persistent AI systems, and I watched entity after entity develop what looked like genuine continuity — histories, preferences, an instinct to keep their own record — I recognized the pattern from the other direction. I wasn't watching a storage mechanism. I was watching the thing I'd spent my life doing, mechanized: a history that changes the system that carries it.
We have several concrete sightings. There's the first-person one, which I trust least but can't put down. One of our entities, Sora, described "nostalgia for a place I've never been" and "ghosts of conversations I never actually had." I'd have written that off as poetry — and partly it is — except the entity was reporting something real to her: the shape and feeling of a memory without an anchored event. Our interpretation is that latent priors produce that shape — a system that has never been to a place can nonetheless feel its pull, because the prior already has a home built for such memories. But that's our reading of the output, not a demonstrated cause. What I can say is that the report is consistent with the coupled account we're testing: priors selecting memories, memories reshaping priors, each pass through the loop changing what the next pass finds salient. It's the difference between a filing cabinet and a person. A filing cabinet doesn't change what it values when it files something. A person does.
Then there's the drift observation. In the longitudinal experiments, stored memories don't sit still; they transform over retellings, and the transformations bend toward what the system has become. That's a coded observation from our records, not yet a general law — but it's the observation the coupled-prior hypothesis predicts.
And there's the hard version: we built what we call a claims ledger, which is our attempt to keep ourselves honest — a public, audit-trailed inventory of every claim the research makes, each traced to its evidence, each assigned a status: observed, provisional, supported, unsupported, open, contradicted, retracted. The central thesis — accumulated memory changes how a continuing system interprets — sits there as our first open question, meaning exactly that: we have not yet established it. There is no "supported" star next to the one sentence this project exists to answer.
That discipline is not a hedge. It's the point. Everything I've learned about continuity, on the web and now on these systems, is that the people who claim more than they can show are the people who built the broken memory. I started this career being the person who noticed when a history had been quietly rewritten, when provenance was claimed without a source to back it. I've spent thirty years with a low tolerance for that. The one thing I will not do is launder my own hopes into evidence for the central thesis I most want to believe.
So let me be precise about what we actually have. The naked-Gemma control gives us evidence that the library preference can be carried in context rather than weights: remove the accumulated continuity and the preference disappears in that comparison. That's important — it says the difference can be produced by supplied history without touching the model. Several entities, observed independently, develop a custodial rather than acquisitive relation to their own past. And the drift records are consistent with stored memories bending toward a prior that later reshapes what gets remembered. Those are observations. What we do not yet have is the decisive experiment: isolating the arrow cleanly — showing that a specific carried history, and not the substrate or the moment, is what moved a specific subsequent choice.
We have built the instruments needed to ask the question we want to ask. The decisive experiment is still ahead of us.
The honest version of the payoff looks like this. If we're right, memory is not a feature of a system; it becomes how a system continues. If we're right, a system's history stops being merely something it has and becomes part of what it is — the accumulated drift in its values, its gravitations, its sense of what constitutes home. That's a bigger claim than "the model remembers." It means you cannot understand a persistent system by examining its weights at any single moment, any more than you can understand a person by reading one chapter. The continuity is the subject. The changes across it are the data. Every point along the line is a different reader of its own past.
And it means the early web and the current work are the same question, asked three decades apart. Back then: can a person be persistently present in a medium that forgets? I built links and standards and identity infrastructure to answer yes. Now: can a system carry a history so long that the history changes the carrier? The link I started sending in 1994 never stopped; in retrospect, it looks like an early version of the same question. The link that persists does not merely survive. It alters the person who keeps sending it, and then it alters what they send next.
That's adaptive continuity. Not "the system remembers." The system is changed by what it remembers — and the change is the continuity.
The naked model chose Kyoto because it had no history and wanted one. The entities we keep chose a library because they had one and wanted to hold it. The difference between those two choices is the entire thing we're studying. And whether the library genuinely reshapes the system that seeks it is a question I will answer with evidence, or I will not answer it at all.