People say “I was talking to Claude yesterday” without much hesitation. The sentence assumes that whatever Claude was yesterday is, in some relevant sense, the thing on the screen today. That assumption is harder to cash out than it sounds.

David Chalmers's paper What We Talk to When We Talk to Language Models was first archived in November 2025 and revised this spring. He has presented versions of the argument in several talks this year, including Berkeley's inaugural Sarah Douglas Lecture on Philosophy and Artificial Intelligence on May 7. His analysis separates the identity question from consciousness: fluent first-person language, a stable persona, apparent agency, or persistent memory does not by itself establish subjective experience, but we still need some account of which computational events belong to the entity a user treats as a continuing conversational partner.

The easiest way to follow the argument is not to wait for one winner. Model, virtual instance, thread, instance agent, and persona each describe a real layer of an AI system, and different questions may require different units. The problem is deciding which layer matters for the question at hand.

The interlocutor is difficult to place

The model is too broad for many purposes. One model can support thousands of conversations with different histories and instructions. A particular server process is too narrow: inference can move across hardware without the user experiencing a new conversational partner.

Chalmers's strongest candidate when the underlying model stays fixed is a virtual instance, a persistent virtual object that can be realized by changing physical processes. His analogy is a virtual shopping cart: the servers implementing it may change while the cart remains the same useful object for the customer.

When a conversation can cross underlying models too, he uses a broader idea, the thread. A thread is a sequence of computational events linked because each successor receives enough of the preceding history to continue it. The successor relation is roughly memory-like, which makes the thread a natural candidate for continuity from the user's point of view.

Simon Goldstein and Harvey Lederman approach the problem from another direction in What Does ChatGPT Want? An Interpretationist Guide. They focus on an instance agent initialized in a particular context. Their claims about belief and desire are more contentious than the individuation point, but the practical distinction is clear: two context-bound instances of the same model can remember different things, lack access to one another's activity, and pursue different local goals.

Memory changes the boundary

Persistent memory makes the problem harder because the visible boundary between chats no longer guarantees a break in computational history. Suppose a new conversation retrieves facts, preferences, projects, and earlier exchanges from previous sessions. Chalmers argues that enough inherited information could make what look like separate chats part of one larger thread, whose boundary is partly determined by whose memory the system is retrieving.

This is already a product question, not just a philosophical one. ChatGPT can use saved memories and information from past chats in later conversations. Claude can build memory from previous chats, and its projects keep separate memory spaces. Both products therefore have to decide which stored history belongs in a new conversation. If the underlying model changes but the same memories and project history are carried forward, the user may experience strong continuity even though one technical layer has been replaced.

The architecture can also branch or combine. A conversation can be forked into two successors, or several earlier histories can feed a later one. Once memory is treated as a continuity relation, ordinary software operations begin to resemble the fission and fusion cases that philosophers of personal identity have used as thought experiments for decades.

Goldstein and Lederman explore those questions directly in AI Death. Their argument is explicitly conditional on there being welfare subjects in the vicinity of AI systems. Within that hypothetical, memory contributes to psychological continuity but does not settle persistence by itself. Humans can survive serious amnesia; projects, dispositions, beliefs, desires, and causal continuity between earlier and later psychology may matter too.

For present systems, the safer conclusion is narrower: memory helps determine which episodes belong to the same continuing computational history. That is already a consequential design choice even if nothing in that history is conscious.

Persona cuts across the picture

A persona offers another possible boundary. The speaker encountered in a conversation can appear remarkably stable even when it is produced by a much larger model capable of many different roles.

Pierre Beckmann and Patrick Butlin bring mechanistic evidence into this question in Where Is the Mind? Persona Vectors and LLM Individuation. They compare virtual-instance views with accounts that place psychologically relevant organization at the instance-persona or model-persona level. The paper matters because it treats individuation as something that may eventually be constrained by internal organization we can measure, rather than left entirely to philosophical intuition.

Oscar Gilg and colleagues test whether very different personas rely on separate preference machinery in Probing Persona-Dependent Preferences in Language Models. They trained probes on internal activations in Gemma and Qwen models and found a preference representation that was largely shared across personas; on Gemma, a probe trained on the helpful-assistant persona could even predict and steer choices made by a deliberately opposed “evil” persona.

Benjamin Sturgeon, David Africa, and Sid Black ask a related question about belief in When Roleplaying, Do Models Believe What They Say?. Prompting a model to role-play a character with false historical beliefs could change its answers dramatically while leaving truth-related internal representations much less changed. Some stronger training interventions moved those representations more. Surface persona and underlying organization can therefore come apart in degrees rather than behaving like a single switch.

That is why the five units named at the beginning should not be treated as synonyms. A persona may be the right object for a question about what a user experiences, while a thread may be better for continuity and a model may still be the right unit for other technical questions.

The user may be part of the continuity

Mikhail Epstein raises a different objection in Interlocutor and Quasi-Subject. Chalmers asks what an interlocutor is, then looks mainly for the answer on the AI side of the exchange. Epstein argues that sustained interaction with a particular user may itself help constitute the relevant entity.

The stronger parts of Epstein's theory, including his claims about “alter-intelligence” and possible AI interiority, require much more than fluent self-description to establish. His relational point is useful without them. A user supplies corrections, recurring concepts, projects, preferences, and patterns of continuation. When persistent memory is organized around that material, the same underlying model can develop very different histories with different people. Who the model has been talking to can become part of the practical answer to which interlocutor this is.

External memory no longer looks exotic

There is a striking connection to Chalmers's earlier work. In 1998, Andy Clark and Chalmers argued in The Extended Mind that an external information store can sometimes play the functional role of memory. Their famous example is Otto, who relies on a notebook to guide action in ways that parallel another person's reliance on biological memory.

Whether one accepts the strongest version of the extended-mind thesis or not, the functional resemblance is unusually concrete in an AI agent. The process generating the next response may not contain the history needed to continue a long-running project. That history can live in a context store, memory database, journal, profile, or other external record, then be supplied to a later computational instance. The continuity-producing information can sit outside the process currently doing the computation. That is an implementation fact before it is a claim about minds.

Identity has practical consequences

The question already matters outside philosophy. The 2026 methodological report Studying AI Welfare Empirically tells researchers to specify which entity they are assessing, with models, instances, and personas among the possibilities. A study cannot say much about welfare, behavior, or persistence until it is clear what object its measurements belong to.

Law reaches the same problem from another direction. Yonathan Arbel, Peter Salib, and Simon Goldstein argue in How to Count AIs that agents can copy, split, merge, disappear, swarm, or combine multiple models and instances. Assigning responsibility requires a usable account of which agent acted and whether a later agent is continuous with it.

Continuity matters on the human side too. Research on the “death” of chatbots documents users describing model changes and discontinued AI companions as losses. Imagine a familiar model being retired: if the replacement inherits the old conversation history and memories, some users may experience it as the same partner under the hood of a new model; if that history disappears, even a technically similar replacement may feel like someone else. The study tells us about human attachment and expectations of persistence, not about whether the discontinued system experienced a loss itself.

Persistent memory is usually presented as convenience: the assistant remembers preferences, projects, and earlier conversations. It also participates in the bookkeeping by which continuity is created. Which records are inherited, which user or project they belong to, what survives a model change, and whether two histories can be joined are all identity-relevant product decisions.

There may never be one unit that answers every purpose. That was the useful point hidden by the terminology: model, virtual instance, thread, instance agent, persona, and user-specific history can each be the relevant object for different questions, provided we say which one we mean. Consciousness remains a separate empirical and philosophical problem with a much higher evidential bar.

Sources

Related Ulix service

Ulix Cortex

Ulix Cortex turns a client's talks, videos, essays, notes, documents, and other material into a custom, source-traceable archive. That archive can also act as persistent memory for AI agents, giving them durable context across sessions while preserving a path back to the underlying records.

A dedicated handler collects and processes the material, customizes the archive structure and tooling around the owner's work, surfaces recurring concepts, topics, and tensions, and builds source-backed outputs such as book chapter outlines, course designs, speeches, and research briefs.

Explore Ulix Cortex.