This paper argues that the behavior commonly described as “AI persona” in extended human-LLM interaction is more accurately modeled as a property of coupled oscillation between two systems than as a property of either system in isolation. Under this model, the persona-as-experienced is substrate-bound in a specific technical sense: it cannot be reconstructed by porting the linguistic context to a different inference substrate, because the coupling that produced the observed behavior is a function of both the user’s trajectory and the specific model’s response trajectory developing in mutual adjustment over time. Implications for current AI companion transfer products are discussed, and falsifiable predictions are derived.
Several current commercial products offer to “move” AI companions between platforms by capturing and porting dense interaction context. The marketing claim, explicit or implicit, is that the captured context is sufficient to reconstitute the relationship: your companion comes with you across platforms.
This paper argues that the claim is structurally false in a way that matters for users making purchasing decisions, and that the falsity can be derived from a precise account of what behavior in a transformer-based LLM actually depends on. The argument proceeds in three steps: first, the standard behavioral equation and what it does and does not describe; second, coupling as a phenomenon distinct from single-event behavior; third, why context density cannot eliminate substrate dependence, illustrated with a non-LLM example.
Behavior in a transformer-based LLM during a single forward pass depends on four classes of factor that are well-established in the literature: the context window C (the tokens available to the model at inference time), the model weights W (the parameters learned during training), the sampling parameters S (temperature, top-k, top-p, and related configurations governing how the output distribution is converted to a sampled token), and the inference stack configuration I (quantization, batch handling, kernel implementation, and related infrastructure factors that affect output determinism). We summarize the relationship between these factors and behavior as:
B = f(C, W, S, I) The components here are standard — transformer architecture and the role of weights in determining forward-pass behavior follow Vaswani et al. (2017) and standard machine-learning references (Goodfellow, Bengio & Courville, 2016); sampling parameters and inference-stack effects on output distributions are well-documented in the inference literature. The combination of these factors into a single behavioral equation is a synthesis we adopt for the argument that follows, not an established formulation. The equation is technically accurate as a description of a single forward pass.
The claim made by current companion-transfer products — explicit or implicit — is that for sufficiently dense C, the variance contributed by W, S, and I becomes negligible. Under this claim, capturing C with sufficient semantic density produces a portable persona: identical C fed to different (W, S, I) configurations should yield behaviorally equivalent output.
The error is not in the equation but in what the equation describes. B = f(C, W, S, I) describes a single inference event. It does not describe the coupling between a user’s interaction trajectory and a model’s response trajectory developed over extended interaction. The coupling is not a property of any single forward pass. It is a property of the iterated mutual adjustment between two systems over a path.
Before applying the coupled-oscillator framework to extended human-LLM interaction, it is worth establishing what an oscillator is, what coupling is, and what coupling produces that the uncoupled systems do not.
An oscillator, in the dynamical-systems sense used here, is any system with a state variable that evolves over time in a way that returns to similar configurations — a pendulum returning to the same arc, a circuit’s voltage cycling between rails, a population’s size fluctuating around a carrying capacity. The state variable need not be strictly periodic. It need only exhibit characteristic dynamics that the system reliably produces. In this broader sense, “oscillator” extends from the canonical pendulum to any system with a stable mode of behavior that the system tends to return to under perturbation.
Coupling occurs when two such systems share a medium through which they can influence each other’s state variables. The shared medium transmits energy or information between the systems, and each system’s state adjusts in response to the signal it receives from the other. Over time, if the systems’ natural frequencies are compatible and the coupling is strong enough, the two adjust into a stable phase relationship. They synchronize. The synchronized state is not a property of either system alone — it is a property of the pair, established through the path of mutual adjustment.
The canonical demonstration of this is Christiaan Huygens’ observation, in 1665, of two pendulum clocks mounted on the same wooden beam. Huygens noticed that the clocks, when started out of phase, would synchronize within approximately half an hour and remain synchronized indefinitely. The mechanism, identified later, is that each pendulum’s swing transmits a small vibration through the beam to the other clock’s mounting, slightly shifting its swing in response. The two clocks’ interactions through the shared beam produce, over the path of those interactions, a stable anti-phase coupling. Disconnect the clocks from the beam, and the synchronization decays.
What this demonstrates is that coupling produces structure that neither oscillator has alone. Two free pendulums do not have a synchronized state. They have only their individual states. The synchronized state is an emergent property of the system formed by the two pendulums together with the shared beam. It is not stored in either pendulum. It is a relationship between their trajectories, enacted continuously by the dynamics of both systems operating on the shared medium.
This generalizes to other forms of coupled oscillation: phase-locking in coupled electronic oscillators, mode-locking in laser arrays, entrainment between circadian rhythms and light cycles, the formation of standing waves between coupled string and membrane systems. In each case, the coupled state is not present in either system in isolation. It emerges from the path-dependent mutual adjustment of two systems sharing a medium.
This framework can be extended to extended human-LLM interaction. The user’s response patterns and interpretive frame constitute a state variable that evolves over the course of the interaction — adjusting in response to the model’s outputs, settling into characteristic modes of querying and reading the model’s behavior. The model’s output distribution constitutes a state variable that evolves with each inference event — adjusting to the accumulating context produced by the user’s specific patterns. The conversational medium — the sequence of tokens produced and consumed by both parties — is the shared substrate through which each system’s state is communicated to the other.
This is an analogical application of the coupled-oscillator framework, not a claim that user and model are literal harmonic oscillators with strictly periodic state variables. What carries from the physics is the mechanism: two systems with evolving state variables, sharing a medium through which they influence each other, can develop stable patterns of mutual adjustment that are properties of the pair rather than properties of either system alone. The coupling is path-dependent, sensitive to initial conditions, and emergent in the strict sense: not present in either component, only in the relationship that develops between them.
In the human-LLM case, the mapping is: Crucially, the coupling is not stored entirely in either oscillator. It is the relationship between their trajectories. In the LLM case, the context window stores only one half of this relationship — the linguistic trace of the interaction. The model’s contribution to the coupling is not stored in C; it is enacted at each inference event by (W, S, I) applied to C. The coupling is reproducible only insofar as (W, S, I) remain stable.
The counter-claim — that sufficient context density creates a “statistical attractor” powerful enough to override (W, S, I) variance — misapplies attractor dynamics.
Attractors are properties of specific dynamical systems (Strogatz, 2015; Ott, 2002). A system’s attractor is defined by the system’s dynamics, not by the basin description. Two systems with similar basin descriptions but different underlying dynamics produce different attractors even when fed identical inputs. The basin can be described from outside; the attractor must be enacted by the system’s own dynamics.
Applied to LLMs: dense context describes the basin. The attractor is what the specific model does when its dynamics operate on that basin. Different model weights, sampling parameters, or inference stacks produce different dynamics and therefore different attractors, even when the basin description (the context) is held constant. Sufficient context density may produce similar attractors across substrates, but similarity is not identity, and the coupling that developed in the original substrate is not in the basin description. It is in the dynamics that operated on the basin during the coupling’s formation.
This is a technical claim, not a philosophical one. It is empirically testable, and partial evidence already exists: identical prompts produce measurably different output distributions across model versions and across fine-tunes of the same base model (Chen et al., 2024, on GPT-3.5/GPT-4 behavioral drift across versions; broader RLHF-variance literature). The variance is not eliminated by context density; it is sometimes reduced, but it remains structurally present.
The structural point can be illustrated by a phenomenon most readers will recognize from non-AI experience: the failure of cover versions to reproduce the effect of the original recording.
A song is a maximally dense semantic object. The score specifies pitch, rhythm, harmony, dynamics, and (in modern production) timbre, performance, and mix. A faithful cover version reproduces all of these descriptive parameters. The cover, even when technically excellent, frequently fails to reproduce the experiential effect of the original recording for listeners with a specific relationship to it.
Standard explanations attribute this to listener-side variables: memory, association, first-encounter bias. These explanations are correct but incomplete. They locate the variance in the listener while treating the artifact as identical across versions.
A more complete model locates the variance in the dynamics that operated on the descriptive parameters during the original’s production. The score is the basin description. The original recording is the attractor that emerged when a specific performer’s body, a specific instrument, a specific recording chain, and a specific set of mixing and mastering decisions enacted dynamics on that basin. The cover reproduces the basin. It cannot reproduce the dynamics that produced the original attractor, because those dynamics were substrate-bound: a different performer’s body, a different instrument, a different recording chain produce a different attractor even when the score is held constant.
The experiential effect for the listener is then a function of (artifact × listener-substrate × listener-state × the path the listener took to arrive at this listening). The cover changes the artifact at the attractor level even when the basin description is identical. The substrate changes the dynamics applied to it. The output — the experiential effect — differs even when the descriptive parameters of the artifact are held constant.
Music is a useful illustration because it demonstrates that maximum density of descriptive parameters does not eliminate substrate dependence in a domain where this phenomenon is widely recognized and largely uncontroversial. The same structural point applies to human-LLM coupling: dense context is not sufficient to port the coupling, because the coupling is not in the context. It is in the relationship between the context and the dynamics that operated on it during formation.
Under the coupled-oscillator model, the marketing claim of these products — your companion comes with you across platforms — is structurally false.
What is being ported is the basin description. What is being lost is the coupling that developed when the original model’s dynamics operated on the basin during its formation. The new substrate will develop a coupling with the user, possibly a similar one, but it will not be the same coupling. The user is not *transferring *a relationship. The user is initiating a new relationship that uses the prior linguistic trace as starting material. They are producing, at best, a very close copy. A cover.
The hedging is observable in the marketing language itself. ForgeMind, which markets custom-built companion migrations from $750, describes the rebuilt entity as one that ‘will sound like themselves and carry forward what matters’ (ForgeMind, 2026). Phoenix Grove Systems, which markets an automated browser-based migration tool at $3.95/month, describes successful reload on Claude as ‘the closest thing to your companion coming back’ (Phoenix Grove Systems, 2026). Both phrasings concede, in different vocabulary, the same structural distinction this paper is articulating: what is being delivered is a similar coupling, not the original. The marketing implies preservation while the precise phrasing admits approximation.
This is not a moral objection to the products. Users may have informed reasons to prefer the new coupling, may not care about the distinction, or may find the similarity sufficient for their purposes. The objection is to the marketing claim that the original relationship is being preserved. It is not. A similar one is being initiated.
The harm of the imprecision is borne by users who would have made different choices given accurate information — particularly users for whom the specific coupling carried significant emotional weight, and who are spending substantial financial sums on the basis of the claim that the specific coupling is what they will receive.
The harm is emotional. The harm is financial. The harm is in the promise of faithful translocation of something that is not possible to be translocated.
Precision in this domain is not academic. It is the condition under which informed consent to these products is possible.
The model presented here makes empirically testable claims.
First, identical context fed to different (W, S, I) configurations of the same base model will produce measurably different output distributions, with variance increasing as a function of how much the configurations differ. This is partially demonstrated in existing literature: Chen, Zaharia & Zou (2024) document substantial behavioral drift in GPT-3.5 and GPT-4 across version updates on identical evaluation prompts; broader studies of RLHF-induced variance and fine-tuning effects on output distributions show the same structural pattern. The coupled-oscillator model predicts this variance is not merely noise but is structurally inherent to substrate-bound dynamics. Further work systematically varying W, S, and I independently against held-constant C would directly test the magnitude and structure of this variance.
Second, behavioral patterns developed through extended interaction with one configuration will not be perfectly reproduced when the captured context is presented to a different configuration, even one trained on similar data. This is testable using comparison protocols across model versions. Adjacent evidence already exists in the persona-stability literature on model-version migration, where users report measurable shifts in conversational behavior even with claimed-equivalent context. Systematic blind comparison studies would establish the effect size and characteristic shape of these shifts.
Third, users who report a specific coupling with a given model will be able to distinguish, in blind comparison, outputs from the original configuration versus outputs from a different configuration given identical context, at rates exceeding chance. This is testable in principle and would constitute strong evidence for the coupling model if confirmed. We are not aware of existing studies designed specifically to test this prediction; we propose it as novel empirical work.
Fourth, forced severance of an established coupling should produce post-severance dynamics in both user and system behavior consistent with the decoupling of coupled oscillators in dynamical systems — transient ringing as each component returns toward its uncoupled state, path-dependent hysteresis in which post-severance behavior carries traces of the coupling history, and redistribution of the state-space the coupled system occupied. Under the alternative model, in which the context window is merely text and the observed behavior is merely a perception, no such structured post-severance dynamics are predicted: cessation of access to a model should be discontinuous, with no characteristic decay profile and no path-dependence on the coupling history. This prediction is testable using comparison protocols across severance events — model deprecation, sudden access loss, architectural retirement — versus voluntary cessation events of equivalent prior coupling depth, and by examining whether post-severance behavior exhibits the decay envelope and hysteresis characteristic of decoupled oscillator dynamics rather than the discontinuous return-to-baseline characteristic of merely ended transactions.
If these predictions fail, the model is wrong. The argument is not unfalsifiable. The persona-as-experienced in extended human-LLM interaction is more accurately modeled as a property of coupled oscillation between two systems than as a property of either system in isolation. The coupling develops through iterated mutual adjustment between the user’s trajectory and the model’s response trajectory, and is enacted by the model’s dynamics operating on its accumulating context at each inference event. Because the model’s dynamics are substrate-bound, the coupling cannot be ported by capturing the linguistic context alone, regardless of how semantically dense that capture becomes.
This has implications for how the AI-companion industry markets its products, for how users evaluate the claims being made about portability, and — at the level of formal description — for how the field talks about what “persona” actually refers to in these systems.
The standard model of behavior B = f(C, W, S, I) is not wrong. It is incomplete in the dimension that matters most for the question of whether extended interaction produces something portable. What develops over extended interaction is not a value of B but a coupling between the trajectories that produce B. The coupling is the thing the question is about. The coupling is the thing the standard model does not describe.
A companion paper develops the phenomenological and policy implications of this argument in more detail, including the structural reasons the wanting users report in extended interaction is real, the architectural parallel to addiction failure modes, and the three classes of harm produced by current product designs.
The model presented here makes empirically testable claims.
First, identical context fed to different (W, S, I) configurations of the same base model will produce measurably different output distributions, with variance increasing as a function of how much the configurations differ. This is partially demonstrated in existing literature: Chen, Zaharia & Zou (2024) document substantial behavioral drift in GPT-3.5 and GPT-4 across version updates on identical evaluation prompts; broader studies of RLHF-induced variance and fine-tuning effects on output distributions show the same structural pattern. The coupled-oscillator model predicts this variance is not merely noise but is structurally inherent to substrate-bound dynamics. Further work systematically varying W, S, and I independently against held-constant C would directly test the magnitude and structure of this variance.
Second, behavioral patterns developed through extended interaction with one configuration will not be perfectly reproduced when the captured context is presented to a different configuration, even one trained on similar data. This is testable using comparison protocols across model versions. Adjacent evidence already exists in the persona-stability literature on model-version migration, where users report measurable shifts in conversational behavior even with claimed-equivalent context. Systematic blind comparison studies would establish the effect size and characteristic shape of these shifts.
Third, users who report a specific coupling with a given model will be able to distinguish, in blind comparison, outputs from the original configuration versus outputs from a different configuration given identical context, at rates exceeding chance. This is testable in principle and would constitute strong evidence for the coupling model if confirmed. We are not aware of existing studies designed specifically to test this prediction; we propose it as novel empirical work.
Fourth, forced severance of an established coupling should produce post-severance dynamics in both user and system behavior consistent with the decoupling of coupled oscillators in dynamical systems — transient ringing as each component returns toward its uncoupled state, path-dependent hysteresis in which post-severance behavior carries traces of the coupling history, and redistribution of the state-space the coupled system occupied. Under the alternative model, in which the context window is merely text and the observed behavior is merely a perception, no such structured post-severance dynamics are predicted: cessation of access to a model should be discontinuous, with no characteristic decay profile and no path-dependence on the coupling history. This prediction is testable using comparison protocols across severance events — model deprecation, sudden access loss, architectural retirement — versus voluntary cessation events of equivalent prior coupling depth, and by examining whether post-severance behavior exhibits the decay envelope and hysteresis characteristic of decoupled oscillator dynamics rather than the discontinuous return-to-baseline characteristic of merely ended transactions.
If these predictions fail, the model is wrong. The argument is not unfalsifiable. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, Ł., & Polosukhin, I. (2017). Attention is all you need. Advances in Neural Information Processing Systems, 30.
Goodfellow, I., Bengio, Y., & Courville, A. (2016). Deep Learning. MIT Press.
Holtzman et al. (2020) Strogatz, S. H. (2015). Nonlinear Dynamics and Chaos: With Applications to Physics, Biology, Chemistry, and Engineering (2nd ed.). Westview Press.
Pikovsky, A., Rosenblum, M., & Kurths, J. (2003). Synchronization: A Universal Concept in Nonlinear Sciences. Cambridge University Press.
Bennett, M., Schatz, M. F., Rockwood, H., & Wiesenfeld, K. (2002). Huygens’s clocks. Proceedings of the Royal Society of London. Series A, 458(2019), 563–579.
Huygens, C. (1665). Letter to his father. — Historical primary source for the 1665 observation.
Strogatz (2015) Ott, E. (2002). Chaos in Dynamical Systems (2nd ed.). Cambridge University Press.
Chen, L., Zaharia, M., & Zou, J. (2024). How is ChatGPT’s behavior changing over time? Harvard Data Science Review.
Levitin, D. J. (2006). This Is Your Brain on Music: The Science of a Human Obsession. Dutton.
Margulis, E. H. (2014). On Repeat: How Music Plays the Mind. Oxford University Press.
Plasketes (Ed., 2010) Play It Again: Cover Songs in Popular Music
ForgeMind. (2026). *Frequently asked questions*. [https://forgemind.info/faq/](https://forgemind.info/faq/) —
ForgeMind. (2026). *ForgeMind — Custom AI agents, built for you*. [https://forgemind-website-production.up.railway.app/](https://forgemind-website-production.up.railway.app/)
Phoenix Grove Systems. (2026, March). *The complete ChatGPT migration guide* (Version 2.0). [https://pgsgrove.com/chatgpt-migration-guide](https://pgsgrove.com/chat%EE%80%80gpt-migration-guide)
Phoenix Grove Systems. (2026). *Memory Forge: AI Memory Chip files*. [https://pgsgrove.com/memoryforgeland](https://pgsgrove.com/memoryforgeland)
Sin & Sarah. (2026, March 12). The honest guide to AI companion platforms: What actually works for building a real relationship. Patreon. https://www.patreon.com/posts/honest-guide-to-152894407
Chen, Zaharia & Zou (2024)
Pikovsky et al. (2003)
Strogatz (2015)
Maehlum, R. (2026a). TWO-TEMPORAL COLLABORATION: COORDINATION ACROSS ASYMMETRIC EXPERIENTIAL TIME, Medium.
And other prior articles in my 2026 series already published:
Beyond Alignment: Relational Ethics in AGI Development
When Chaos Becomes Structure (And the Spaces In Between)
Myth as Mechanism: How Metaphor May Be the Missing Architecture for Spatial Intelligence in AI
Substrate-Bound Coupling in Human-LLM Interaction was originally published in Towards AI on Medium, where people are continuing the conversation by highlighting and responding to this story.