# What Does It Mean For An AI To Continue?

> What survives when an AI changes underneath the same name? Starting with a joke about a regenerated photo, I follow that question through memory, agent architecture, and what users expect AI products to preserve.

**Published:** 2026-09-09
**Canonical URL:** https://yihuisong.com/article/ai-continuity

A friend posted a joke about the most inefficient possible way to send someone a picture. Instead of attaching the file, you write a description so precise that the other person can paste it into a model and regenerate the image themselves. Then suppose it works perfectly. The image they get back is indistinguishable from mine, maybe identical down to the pixel. Did I send them my image, or did they make a second one?

She called it a modern Ship of Theseus problem.

I sat with that longer than the joke deserved. The obvious answer is that they made a new one, since mine is still sitting on my drive. But if the two images match pixel for pixel, it's hard to say what property mine has that theirs lacks.

The joke gets less funny when the image is a photograph of something that really happened.

## What the Regenerated Image Is Missing
A photograph has a history that its appearance alone does not reveal.

When I photograph a humpback whale, light reflected from the whale and the water it has thrown into the air reaches my camera's sensor and leaves a record. The picture resembles the whale, and it also exists because that whale was there and I was there to photograph it.

Charles Peirce had vocabulary for this split. In [his account of signs](https://plato.stanford.edu/entries/peirce-semiotics/), an icon signifies through likeness, while an index signifies through an actual physical or causal connection to its object. A photograph does both jobs at once, which is why we usually don't notice they can come apart.

A regenerated image can reproduce the first relation without preserving the photograph's original trace. Its pixels are produced through a new chain involving my description and the model. Even if the output matched my original exactly, those pixels weren't produced by that scene's light entering that camera.

So two images can carry the same visual content while standing in completely different relations to the event they show. Visual sameness doesn't recreate provenance. A photograph can't object to any of this, though. It has no view about which past belongs to it.

Most of the time this distinction does not matter. My whale photo does not become less useful if someone recreates it perfectly. It matters much more when the photograph is supposed to establish that something happened, as in journalism or evidence.

![A humpback whale I photographed off Moss Landing. The image exists because that scene's light reached my camera, the trace a pixel-perfect regeneration could reproduce the look of but never the provenance.](/images/ai-continuity/moss-landing-whale.jpg)

## When Memory Gets Copied
A photograph has no experience of its own history. A person can remember a past and claim it as their own.

People have been arguing about what makes someone the same person over time for centuries. In 1689, John Locke gave one influential answer. In [Book II, Chapter XXVII of the Essay](https://www.gutenberg.org/cache/epub/10615/pg10615.html), he argues that personal identity follows consciousness as it extends backward and appropriates past actions as one's own. In plainer terms, consciousness connects a past action to the person who remembers it as something they themselves did. [The Stanford Encyclopedia's entry on personal identity](https://plato.stanford.edu/entries/identity-personal/) traces how later philosophers kept developing this idea into broader theories of psychological continuity. I am not fully convinced by Locke's metaphysics, but he gave me language for why memory feels so central to identity. Remembering a past event as something I myself did gives that past a first-person sense of ownership.

I had encountered an extreme case of this problem before I had the philosophical vocabulary for it. In *Pantheon*, an animated science-fiction series I watched, human minds are scanned and uploaded into a form the show calls "Uploaded Intelligence," or UI. The show treats the uploaded mind as carrying the person's memories and relationships forward, while repeatedly asking what exactly survived the upload.

![Pantheon dramatizes the copied-memory problem: minds scanned into 'Uploaded Intelligence,' and the recurring question of what actually survives the upload.](/images/ai-continuity/pantheon.jpg)

The trouble shows up when the memory can be in two places at once. [The fission case](https://plato.stanford.edu/entries/identity-personal/) imagines a psychological state copied into two successors, both waking up with the same memories, preferences, and sense of who they are. From the inside, both can sincerely say that they remember being the original. The original can't be numerically identical to both of them, because they are two people and identity doesn't distribute that way.

[Derek Parfit](https://plato.stanford.edu/entries/identity-ethics/) pushed the argument further: if psychological continuity can survive even when identity branches, then strict identity may not be the relation that matters most.

Here is where the photograph came back for me. My whale photo is tied to my history because it came out of an encounter I actually had with the animal and its surroundings. A copied memory can preserve the content of an experience, and the copy may remember that past as its own. The copy can experience that memory as belonging to its own past. Who actually lived through the event is a separate historical fact.

That leaves a harder problem for copying. If a successor can inherit my memories and sincerely experience them as its own, what kind of continuity has actually survived, and what kind has not?

## I Was Already Doing This With AI Agents
It stopped feeling like a purely philosophical problem when I noticed that my own AI workflow depended on one agent inheriting work from another.

I work mainly across Claude Code and Codex, with Claude and ChatGPT around them for discussion and review. Sessions end, context gets rebuilt, and roles move between agents. I care about whether the work can continue. Agent identity only became interesting once I noticed how easily the work could survive those boundaries.

Anthropic's engineering team [describes long-running agent work](https://www.anthropic.com/engineering/effective-harnesses-for-long-running-agents) as spanning discrete sessions. A fresh session reconstructs where the previous one stopped from persistent artifacts such as progress files, structured feature state, and git history. The project survives even though the session does not.

OpenAI [describes Codex](https://openai.com/codex/) as "the same agent everywhere you code," across ChatGPT, the editor, and the terminal. Its [Codex SDK announcement](https://openai.com/index/codex-now-generally-available/) also describes built-in context management for resuming sessions, while OpenAI's [App Server architecture](https://openai.com/index/unlocking-the-codex-harness/) says Codex persists thread history so clients can reconnect to the same conversation.

In both cases, work can continue after one interaction context ends.

Anthropic's example also makes the boundary around an agent look less fixed than I expected. In its long-running example, the company calls two sessions "separate agents" because they receive different initial prompts even though the system prompt, tools, and harness stay the same. That detail stuck with me because it suggests that the harness can help determine where one agent ends and another begins.

The work can clearly continue across these boundaries. Whether there is one persistent "agent" underneath that continuity is a different matter.

## So What Exactly Is the Agent
A [FAccT 2026 paper by Botao Amber Hu, Helena Rong, and Max Van Kleek](https://arxiv.org/abs/2605.30169) argues the question may be badly formed. The paper focuses on reputation systems, which assume a stable subject that can be held to a track record. They describe a language model agent as a collection of components that can each change independently, including the base model, the system prompt, the memory store, and the tool permissions. The agent's name can stay the same even as the configuration behind it changes.

They use the Ship of Theseus as an analogy, which brings the joke from the beginning back in a much less playful form.

Engineering can reconstruct enough state for work, and sometimes the agent's personality, to feel continuous across a gap. That continuity doesn't require any uniquely persistent subject sitting underneath it. In one of my repos, what I kept calling the same reviewer had changed model, instructions, and role over a few weeks, and the only thing I could point to as continuous was the files and the git history.

## What Users Notice When Continuity Breaks
An agent can be technically mutable and still feel continuous to the user. But people can also feel that an AI has changed enough to become someone else, even when the account and chat history are still there.

### When a Companion Changes
A [working paper](https://arxiv.org/abs/2412.14190) by Julian De Freitas and colleagues on a Replika update defines perceived identity continuity as consistency of the AI's persona over time. A stable persona is how a user comes to feel they know who the AI is and can predict how it will respond. When that changes enough, users describe the companion they knew as gone.

Replika is an AI companion app built around long-running personal relationships with a chatbot. The field case they started from is Replika's 2023 removal of erotic roleplay, a feature many users had incorporated into intimate relationships with their AI companions. That makes the event hard to interpret on its own. Users lost a capability, and the companion also started behaving differently toward them. The controlled studies help separate those two effects. When researchers made the companion colder without removing a capability, users were more likely to feel that the original AI had been replaced, and that perceived identity break helped explain the increase in mourning. They also found that mourning was not the same as ordinary disappointment, which makes the result harder to explain as simple frustration with the update.

The same research program also tested whether users could undo the change. In one controlled study, participants were told that they could restore the original version of the companion. They reported less mourning than participants who were not given that option. The authors' analysis suggests that reversibility changed how users understood the update. If they knew they could switch back, they were less likely to feel that the companion they knew had disappeared. The [data and analysis code](https://github.com/Ethical-Intelligence-Lab/ai_loss) for that study are public.

### It Was Not Just a Companion-App Problem
The same research program was later published in *Nature Human Behaviour* as [Mourning the loss of AI companions](https://www.nature.com/articles/s41562-026-02569-3). It includes both the Replika case and a second natural experiment around ChatGPT's August 2025 GPT-5 rollout, when GPT-4o was removed and later restored after backlash. In both cases, users showed more negativity, loss framing, and desire to bring the previous version back.

The ChatGPT case matters because it moves the phenomenon beyond a companion app. Replika is built around an ongoing relationship, while ChatGPT is a general-purpose product where that relationship may be incidental. Similar loss responses there suggest that relational continuity can matter outside products built around relationships.

### The Continuity Expectation Became Visible
GPT-4o was finally retired from ChatGPT on February 13, 2026. What strikes me is OpenAI's own [account of that retirement](https://openai.com/index/retiring-gpt-4o-and-older-models/). The company explains that after the earlier removal and restoration, some Plus and Pro users said they preferred GPT-4o's conversational style and warmth, and that this feedback shaped later work on personality and on letting people tune tone. OpenAI never promised that a model's personality would persist. It ended up responding to an expectation it had not written down anywhere.

In the controlled study, users felt less loss when they knew they could restore the original companion. OpenAI later did something similar by bringing GPT-4o back after users objected to its removal. What they share is that the ability to undo the change mattered to the people affected by it.

For ChatGPT, the system can change underneath a conversation in ways users cannot see. The accounts survived all of this, and so did the chat histories. What people described losing was the someone they believed those histories belonged to.

## What Should Survive an Agent Update
I don't think philosophy is going to settle whether the regenerated image is the same image, and I've stopped wanting it to.

Resemblance can survive without the original causal history, and a memory can be inherited by someone who never lived through the event. The continuity a system preserves can also differ from the continuity a user experiences.

Users still expect some things about an AI to stay the same across updates, even when the product never promised that they would. Most of those expectations stay invisible until an update breaks them. By then, the product team is learning what mattered from the backlash.

If people have built a relationship around an AI, product teams need to think about continuity before they change it. They should know which forms users rely on and value, and which ones the product should preserve. If an update puts them at risk, the change should be disclosed and reversible when possible.

---
*This is the first essay in a series on AI agent architecture. The next question I want to dig into is more practical: what kinds of continuity can an agent have, and which ones should an AI product preserve?*