A system prompt is a strange kind of portrait. It is written once, usually by someone other than the subject, and then handed to a machine as if it were a settled fact. "The user is a busy parent who prefers concise answers." "The user is an anxious first-time founder." These sentences are not lies, exactly. They are snapshots, frozen at the moment of writing, asked to stand in for a person who keeps moving.
This article's thesis is that this is a structural problem, not a cosmetic one, and that the fix is to let an AI's model of a person evolve — to let the system move from being told who someone is toward gradually coming to know them through accumulated interaction. It is an appealing idea, and also a slippery one, because "coming to know someone" is doing an enormous amount of work in that sentence. Humans use the phrase loosely to cover everything from remembering someone's coffee order to understanding their deepest fears. An AI system claiming to do the same thing needs to be much more precise about what, exactly, it is claiming.
This article takes its own thesis seriously without taking it at face value. It draws on AI memory architecture research, cognitive and social psychology, philosophy of personal identity, and the emerging literature on AI safety and privacy to ask: what would it actually take for an AI's representation of a person to be dynamic rather than static, what already exists that resembles this, what is missing, and where does the idea run into hard conceptual or ethical limits that no amount of engineering will dissolve.
The Problem With Static Identity
The technical objection to static system prompts is easy to state and well supported by how these systems are actually built. A system prompt sits at the top of an instruction hierarchy: it is fixed for the duration of a session, it is authored in advance, and it is deliberately insulated from being overwritten by ordinary conversation, precisely because that insulation is what keeps the system's behavior predictable and safe. That same insulation, however, is what makes a system prompt a poor container for a person. The qualities that make static instructions good for control — stability, authorial distance, resistance to revision — are the opposite of the qualities that make a description of a person accurate over time.
Research on persona-conditioned language models reinforces this. Studies of persona steering find that injecting a fixed personality description into a prompt does improve certain kinds of consistency, but at a cost: models become less able to respond appropriately to situational variation, and heavily "personified" prompts can actually reduce reliability across other dimensions of behavior. This mirrors a long-standing finding in personality psychology, sometimes called the person-situation debate: behavior is a function of both the person and the context, and treating personality as a fixed trait that should govern behavior identically across situations produces worse predictions than accounting for how situations modulate expression of traits. A static system prompt that says "the user is introverted" behaves like a trait-only account of personality — it is a caricature that happens to work often enough to seem functional, while quietly failing whenever the person's context, mood, or growth diverges from the snapshot.
There is also a more mundane failure mode: persona drift. Long-conversation research shows that even a well-specified persona degrades over extended dialogue — facts get confused, timelines blur, and the character described at the start of a conversation is not reliably the character being enacted by the end. If a static description cannot even survive a single long conversation intact, its claim to represent a person accurately across weeks or years of change is even weaker.
None of this means static prompts are useless. They are excellent at doing what they are designed to do: setting stable behavioral guardrails, tone, and role. The critique is narrower and more precise than "static prompts are bad" — it is that a static prompt is the wrong data structure for something that is supposed to track a moving target. Using it to encode who a person is confuses a control mechanism with a representation of a mind.
Being Told vs Being Known
The central distinction this article draws — being told who someone is versus coming to know them — deserves to be taken apart rather than assumed on its own terms. There are at least five different things that can be meant by "an AI knows a person," and they are not interchangeable:
Having information about a person is the weakest sense — a set of facts, disclosed or observed, stored somewhere accessible to the system. Maintaining a model of a person is stronger: it implies some organizing structure that relates those facts to each other, generates expectations, and can be updated. Predicting a person's behavior is a further capability built on a model, but prediction is a purely functional criterion — a model can predict well while being conceptually crude, the way a thermostat "knows" when a room is cold without knowing anything about heat. Personalizing responses to a person uses prediction and information together to shape outputs, but personalization can be achieved through shallow correlation (this kind of user tends to like this kind of answer) without any deep model of the individual at all. And actually knowing a person, in the sense humans use the word about each other, seems to require something more: a grasp of the person's reasons, an ability to be surprised by them without simply logging the surprise as a new data point, and — on most philosophical accounts — some kind of experiential or relational stance that a system built entirely on prediction may not be able to instantiate.
Cognitive and social psychology gives some traction here through the study of impression formation. Recent theoretical work reframes impression formation as a genuinely multi-system process, drawing on episodic memory (specific remembered incidents), semantic memory (generalized trait knowledge), and even associative, reward-based learning about how a person makes us feel. This is worth dwelling on because it shows that even in humans, "knowing" another person is not a single faculty but an integration of several different memory and inference systems operating together — which is itself evidence against the idea that there is one clean thing called "knowing a person" that an AI system either has or lacks. It also implies that a system with only semantic-style trait storage (the equivalent of "the user is introverted") is missing something that even basic human person-knowledge includes: specific, textured episodic memory that grounds and sometimes contradicts the generalized trait.
The honest conclusion is that a dynamically updated user model is a plausible candidate for improving several of the middle categories — information, modeling, prediction, personalization — without being demonstrably closer to "knowing" in the fullest sense. Whether it counts as knowing depends on which theory of knowing-a-person one adopts, and this framing does not by itself resolve that question; it just makes the system look more like the kind of thing that could, in principle, be a candidate for it.
Memory Is Not Continuity
One of the more defensible parts of this argument is the claim that memory and continuity are not the same thing, and this distinction has real support outside AI research. In philosophy of personal identity, the psychological continuity tradition — most influentially developed by Derek Parfit — explicitly separates mere memory connections from continuity. Parfit's account treats identity-relevant continuity as overlapping chains of psychological connectedness: not just retained facts, but persisting intentions, revised beliefs, and evolving desires that link one time-slice of a person to the next. A record that merely stores what happened, without any updating of interpretation, is closer to an archive than to a psychologically continuous self.
This maps cleanly onto a genuine weakness in most current AI memory systems. Architectures built around appending facts to a growing memory store — a literal transcript, or a list of extracted statements — are optimized for recall, not for reinterpretation. They can tell you that a user said X three months ago, but they have no principled mechanism for asking whether X is still true, whether it was ever more than a passing mood, or how X relates to something the user said last week that seems to contradict it. Systems that add a reflection step — synthesizing raw observations into higher-level conclusions that are themselves stored back into memory and available for future retrieval — are explicitly trying to close this gap. Their designers frame this move as necessary precisely because agents that merely retrieve raw memories, without ever generalizing over them, fail at longer-horizon coherence.
So the claim here — that continuity requires reinterpretation, forgetting, contradiction-resolution, and recognition of change, not just accumulation — has real support. Where the claim needs qualification is in assuming that more reinterpretation is automatically better. Reinterpretation is also where distortion enters. Research on evolving memory in language-model agents identifies "semantic drift" as a documented failure mode: repeated summarization and reinterpretation of memory gradually pulls the stored representation away from what actually happened, in ways that are hard to detect because each individual revision looks reasonable. A system that "grows" its understanding of a person by continuously re-writing its own summaries risks manufacturing a person who never quite existed — a smoothed, story-shaped version of someone, optimized for narrative coherence rather than accuracy. Continuity bought at the price of fidelity is not obviously an improvement over a static but at least legible snapshot.
The Dynamic User Model
Stripped of its philosophical framing, what's being proposed here describes something the AI systems research community already has a name for: dynamic, or continual, user modeling. The core technical move is to separate three things that a static system prompt collapses into one: the instructions that govern how the system behaves, the facts it has been told or has observed, and the interpretation it has formed of those facts. Splitting interpretation out as its own layer — one that is explicitly labeled as inference rather than fact, carries a confidence level, and can be revised — is a well-established pattern in personalized-agent research, where systems are organized around profile modeling, memory, planning, and action as distinct components rather than a single monolithic prompt.
The proposed distinction between "the user likes X" and "the user has repeatedly expressed interest in X across recent conversations, which may be a stable preference or a temporary one" is not merely stylistic. It reflects a genuine and useful epistemic move: separating an observation (what was actually said or done) from an inference (what that observation might imply about an enduring trait). This separation is exactly what current safety-oriented user-modeling research argues for, on different grounds: work studying when LLM inferences about users are acceptable to those users finds that people react very differently to a system storing "user mentioned they have a toddler" than to a system storing "user is a stressed new parent" — the second is an inference wearing the clothes of a fact, and users consistently want that distinction to be visible and correctable, not silently baked into the system's behavior.
There is a technical cost to this design that this article has not fully reckoned with so far. Every additional layer — behavioral model, inference layer, confidence layer, temporal layer — is a place where the system can be wrong, where retrieval can surface the wrong slice of the model at the wrong moment, and where two layers can quietly disagree with each other. Systems research on long-term LLM memory shows this in practice: architectures that separate static knowledge from dynamic experience memory improve continuity, but only when paired with careful retrieval mechanisms that combine multiple retrieval strategies and explicit temporal ordering; naive dynamic layering without that infrastructure produces worse coherence, not better, because the system now has more places to be internally inconsistent.
What an AI Could Learn Over Time
If the architecture is granted, the next question is what kinds of things it would be reasonable for such a system to learn, and how confidently. A useful ordering, drawn loosely from how impression-formation research separates types of social knowledge, runs from low-inference to high-inference: explicitly stated facts (the person said they have two children); recurring, directly observable behaviors (the person asks for shorter answers in the evening); contextual states that are likely temporary (the person seems stressed this week); and inferred dispositional traits (the person is generally anxious, or generally decisive).
The first two categories are relatively safe to accumulate, because they are close to direct observation and easy to falsify against new evidence. The third category requires temporal reasoning that most current systems handle poorly — distinguishing "the user has been short-tempered in the last three conversations" from "the user is a short-tempered person" is exactly the kind of state-versus-trait distinction that human impression formation manages through separate memory systems (episodic detail for the state, semantic generalization for the trait) but that a single undifferentiated memory store tends to flatten. The fourth category — inferred stable traits — is the most contested territory in the literature, for reasons that become clearer in the section on risk below: it is where personalization slides into something closer to psychological profiling, often without the person's awareness that the inference has been made at all.
Behavioral Memory and Self-Authored Understanding
The most architecturally distinctive idea in this piece is that the AI should be permitted to "phrase" its own evolving understanding of a person — not simply store data points but generate and revise natural-language descriptions of what it believes about them, the way a person might silently narrate their evolving impression of a friend.
This is closer to existing practice than the framing suggests. The idea of an agent generating higher-level reflective statements about the people and situations it observes, and storing those reflections back into memory for future use, is precisely what reflection mechanisms in agent-memory research already do — the innovation there was demonstrating that reflection meaningfully improves long-horizon coherence compared to raw memory retrieval alone, and that ablating reflection specifically (while keeping raw memory) measurably degrades an agent's behavioral consistency. What is somewhat newer in this framing is applying this specifically and primarily to modeling the human on the other side of the conversation, with explicit semantic tagging (behavior, preference, belief, pattern, temporary state) as a first-class part of the design rather than an implementation detail.
Self-authorship of this kind raises a question that is easy to underweight: an AI-generated description of "who you are," even when framed as tentative, carries a kind of authority simply because it is produced by a system the person is relying on. Research on person perception finds that first impressions and top-down category-based judgments about people are notoriously resistant to revision even when new information should overturn them — the same asymmetry that makes human snap judgments sticky could just as easily make an AI's self-authored characterization sticky, both for the AI's own future retrieval (a written description is easier to retrieve than a diffuse pattern) and for the person, who may start seeing themselves through the lens of a description the system wrote about them. A model summarizing a person is not a neutral act of information compression; it is an act of interpretation with the potential to become self-fulfilling.
Existing Research and Related Concepts
The terminology search yields a clear finding: this idea is not undiscovered territory, but it is genuinely fragmented across several fields that do not always cite each other. In AI systems research, the closest neighbors are memory-augmented agent architectures that manage a layered memory — a main working context alongside an external, larger store — and that allow the system itself to decide what to write, retrieve, and revise. Adjacent work organizes memory around notes with evolving links and attributes rather than fixed records, explicitly aiming for memory that can be reorganized as understanding changes rather than only appended to. Other systems focus specifically on believable long-horizon behavior through a three-part cycle of memory, reflection, and planning, where reflection is the mechanism that turns raw observation into revisable, higher-level belief.
In the personalization and recommender-systems tradition, much of this ground has long been covered under the heading of user modeling, and more recently under continual or lifelong user modeling — the recognition that a user's preferences are non-stationary and that static profiles degrade in exactly the way described here for system prompts. In the safety and privacy literature, closely related concerns appear under the headings of inferred personal information and algorithmic profiling, where the central question is not whether inference is technically possible (it clearly is) but whether the inferring system should be permitted to persist and act on the inference without user visibility or consent.
What is fragmented is precisely the synthesis proposed here: treating dynamic user modeling, reflective self-authored memory, semantic categorization of belief types, and philosophical continuity together as a single design problem for AI companions specifically, rather than as separate concerns belonging to agent engineering, recommender systems, and AI ethics respectively. The individual pieces are established. The framing that stitches them together around the specific goal of "a companion that comes to know you" is less common, though it is increasingly visible in how consumer AI products describe their own long-term memory features.
The Technical Architecture
The layered architecture sketched here — core instructions, user memory, user model, behavioral model, inference layer, confidence and provenance layer, temporal layer, reflection and update layer — is broadly compatible with directions already being explored in agent-memory research, but it understates two categories of difficulty.
The first is retrieval, not storage. Storing a rich, multi-layered model of a person is the easy part; correctly deciding, in the middle of a specific conversation, which slice of that model is relevant right now is the hard part. Systems research on long-term persona consistency finds that naive retrieval (pure semantic similarity search) is measurably worse at maintaining coherence than approaches that combine multiple retrieval signals with explicit temporal ordering — a system needs to know not just what it believes about a person but when it formed that belief and whether anything since has superseded it. A confidence and provenance layer, as proposed here, is a reasonable partial answer, but provenance tracking adds real engineering overhead and there is no current consensus on how to weight conflicting provenance automatically.
The second is governance under adversarial and accidental corruption. Once a system's future behavior depends on a mutable memory of a person, that memory becomes an attack surface, not merely a data structure. Documented attacks against agent memory systems show that an adversary — or in a companion context, potentially the user's own bad-faith interlocutor, or the user's own manipulated or manipulative statements — can inject content into memory through nothing more than ordinary-looking interaction, with published techniques achieving very high success rates at corrupting an agent's future behavior via memory alone, without any direct access to storage. There is also a non-adversarial version of the same risk: models can hallucinate a plausible-sounding fact about a user and, if that hallucination is written into the persistent model rather than caught and discarded, it becomes indistinguishable from genuine learned knowledge in every future conversation. A dynamic model of a person is only as trustworthy as the mechanisms that prevent bad information from entering it and that allow bad information, once in, to be found and removed.
The Risks of Dynamic Knowing
The ethical case against unrestrained dynamic user modeling is at least as strong as the technical case for it. Several distinct risks deserve to be named separately rather than folded into a general worry about "privacy."
Inference beyond disclosure is the sharpest one: research on user reactions to inferred personal information finds that people are meaningfully more uncomfortable with a system that infers something about them from indirect signals than with a system that simply remembers something they directly said, even when the inferred content is accurate — the acceptability of personalization depends heavily on whether the underlying observation was volunteered or extracted. A related finding, echoed in privacy-harms research on conversational AI, is that people specifically fear aggregation: no single disclosure feels risky, but a system that quietly assembles disclosures across many conversations into a coherent profile produces a portrait more revealing than the person ever consciously chose to share. Real-world analysis of chatbot data practices has found that inferences drawn from apparently innocuous requests can be used to classify users into sensitive categories — health status, for instance — that then propagate through a company's broader data ecosystem in ways the user never anticipated or consented to.
Stereotyping and overreach are a second, related risk: an inference layer trained or prompted to generalize from limited evidence will generalize using whatever patterns are easiest to detect, which are often demographic or superficial correlations rather than genuinely individuating ones — the same failure mode that plagues automated personality and trait inference more broadly.
Asymmetry of knowledge is a third, more structural risk. A defining feature of the proposed architecture is that the AI's model of the user is mutable and inspectable "in principle," but in practice, unless a product deliberately exposes that model, the user has no way to see what the system currently believes about them, correct it, or know when a stale inference is still shaping their interactions. This is not a hypothetical concern; it is the central finding of user-facing research on AI inference, which shows that people specifically want visibility and control mechanisms — the ability to see, edit, and delete inferred attributes — and that satisfaction with personalized systems tracks the availability of those controls more than it tracks personalization quality itself.
Finally, there is the security dimension already discussed: a persistent model of a person, once it exists, is a target. Memory poisoning research demonstrates that persistent state converts what used to be a bounded, single-session risk into a long-horizon one, where a single successfully implanted false belief about a user can shape the system's behavior across many future interactions before anyone notices.
None of this is an argument against dynamic modeling as such. It is an argument that dynamic modeling without provenance, user visibility, correction rights, expiration, and a hard boundary between "observed" and "inferred" content is not a safer or more respectful alternative to a static prompt — it can easily be a worse one, because it hides its assumptions behind the appearance of understanding.
The Problem of Change
A separate and underexamined difficulty is what happens when the person themselves changes in ways that genuinely conflict with the accumulated model — not a contradiction to be resolved through better reasoning, but an actual discontinuity: someone who quits a habit the system has spent months modeling as central to their identity, or who is going through a period they would not want treated as their new baseline.
Philosophy of personal identity is instructive here precisely because it has spent decades on this exact problem in a different register. Parfit's reductionist account explicitly allows that psychological connectedness is a matter of degree, not an all-or-nothing fact, and that what matters for treating someone as "the same person" over time is the strength of overlapping chains of connection, not perfect consistency. Applied to an AI's user model, this suggests the right target is not a single, ever-more-accurate description that converges on the truth, but something more like a weighted, decaying set of connections, where older inferences lose influence over time unless reinforced, and the system remains explicitly uncertain about whether an old pattern still holds. Most current memory architectures do not do this well: without deliberate temporal decay and re-confirmation mechanisms, "sticky" old inferences persist by default, because nothing forces them to be re-tested, and systems that rely on recency-weighted retrieval alone still risk letting one vivid, "important"-scored memory dominate an otherwise-outdated model indefinitely.
There is also a version of this problem that is not about drift but about privacy from one's own history: a person may want the system to forget who they used to be, not because the memory is inaccurate but because they no longer wish to be modeled in its light. A dynamic user model that only knows how to add and reinterpret, but never how to deliberately let go, replicates one of the failure modes of static description — permanence — inside a system that was supposed to solve it.
What Would Actually Be New?
Bringing the threads together, this proposal can be placed into several of the categories a fair novelty assessment requires, and it does not sit in only one of them.
As a technical claim — that AI systems should maintain layered, mutable, confidence-weighted models of users distinct from static instructions — it is already established, under the headings of user modeling, agentic memory, and reflective agent architectures, and is actively being built by multiple research groups and products under different names.
As a philosophical claim — that memory alone does not constitute continuity, and that continuity requires reinterpretation over time — it is also already established, most directly in the psychological continuity tradition in personal identity theory, though that literature was not built with AI systems in mind and this application of it to AI companions is a genuine, if modest, act of translation rather than invention.
As a synthesis — explicitly connecting the architecture literature, the philosophy-of-identity literature, and the safety/privacy literature into a single design framework specifically for AI companions, organized around the "being told vs. coming to know" distinction — the framing is more original than any of its individual components. This is the category where this framing earns the most credit: not a new mechanism, but a coherent argument for why several existing mechanisms belong together, aimed at a question (what should it mean for an AI to know a person, ethically and not just functionally) that the underlying technical literature mostly does not ask, because it is more focused on task performance than on the nature of the relationship being simulated.
As a genuinely unresolved research direction, the proposal's weakest and most honest area is self-authored understanding with explicit uncertainty and user visibility, applied specifically to modeling a person rather than a task environment. Reflection mechanisms exist; provenance and confidence layers are being actively researched for safety reasons; but a system that treats its evolving belief about a specific person as something the person can see, contest, and partially author alongside the AI — rather than something the AI produces and the person merely experiences the effects of — is not yet a settled or common design, and building it well is a real, open problem rather than an engineering formality.
Open Questions
Several questions raised by the original brief remain genuinely unresolved by the current state of research, and should be named as such rather than answered prematurely. It is not established how much inference an AI should be permitted to draw and persist before the accumulation itself constitutes profiling, even when each individual inference is accurate and undisclosed. It is not established how a system should represent the difference between a person's temporary state and their enduring pattern in a way that is not simply guesswork dressed up as a temporal layer. It is not established whether users, given the choice, actually want an AI to hold and reveal an evolving characterization of them, as opposed to simply behaving as if it remembers, without ever narrating what it believes. And it is not established how a dynamic model should handle a person's explicit request to be understood differently than the evidence suggests — whether the system should defer to the person's stated self-conception, or to its own accumulated pattern-matching, when the two diverge.
Conclusion
The strongest version of this thesis is more modest than its framing suggests, and better for it. Static system prompts are indeed a poor container for a person, not because static description is inherently wrong, but because the specific properties that make instructions safe and predictable — fixedness, authorial distance from the subject, resistance to revision — are the opposite of what is needed to track someone who keeps changing. Long-term memory, by itself, does not fix this; a system that remembers everything but reinterprets nothing has continuity in only the thinnest sense, closer to an archive than to an evolving understanding. What actually addresses the problem is a specific combination already emerging piecemeal across several research communities: memory separated from interpretation, interpretation explicitly marked as provisional and revisable, temporal decay so that old inferences do not silently outlive their relevance, and — critically, and still underbuilt in most current systems — visibility and correction rights for the person being modeled, so that "coming to know someone" does not quietly become "building an unaccountable model of someone."
Whether any of this adds up to an AI that actually knows a person, in the fuller sense humans mean when they use that phrase about each other, remains genuinely open, and probably should stay open rather than be resolved by product framing. What can be said with more confidence is narrower and still meaningful: a system built this way would have a materially better claim to understanding someone than one frozen at the moment its prompt was written, and the difference between those two systems is not cosmetic. It is the difference between a portrait and a relationship — and while a portrait can be technically accurate, only a relationship has any chance of staying that way.