One Model, Many Minds

1. Introduction

Every organisation that puts an AI assistant to work soon finds it needs to play more than one part. For most of their history, large language models worked as general-purpose text predictors. That is changing fast. The systems now being put to work in the real world are expected to be someone in particular: one user's personal assistant, a consistent character in a game, a support agent that speaks in a single steady voice and never breaks role. And whichever part it is playing, the model has to remember which one it is — turn after turn, and often session after session. Serving all of them from one network means several stable identities have to coexist on a single base model without contaminating one another, whether they are written into its weights or supplied to it at run time.

Researchers usually split the ambition into two goals. A 2024 survey names them role-playing, where a model is assigned a persona to play, and personalisation, where it acts on behalf of a user's persona [1].

The answer has two halves, and the rest of this paper follows them in turn — how a persona is written into a model during training, and how it is selected and applied during inference. The distinction sounds procedural. It turns out to shape almost everything about how far the arrangement scales.

This is not just a research question. Most companies that use AI assistants want the same thing: one system that can be a brand voice for customers, a private assistant for staff, and a fixed character in a training tool — instead of building and maintaining a separate one for each job.

2. From a Few Sentences to a Graph of Facts

The field began with something deliberately simple. An early benchmark, published in 2018, handed each speaker a handful of short sentences — "I am a nurse. I love hiking. I have two dogs." — and asked crowdworkers to converse in character from them [2]. It became a widely used test for persona-based dialogue, and a short written profile is still the most common way a persona is handed to a model today. Richer formats followed quickly: a large dialogue dataset whose speakers carry attribute–value traits such as age, gender, location and interest tags, published in 2019 [3], and long multi-session dialogues in which partners refer back to what they learned in earlier sessions, built to test whether models can carry that context across separate conversations, published in 2022 [4].

Plain text is easy to write, and it carries a persona only so far. It does not record how facts relate to one another, gives the model no structure to reason over, and takes more of the context window with every sentence added — precisely as personas become detailed enough to be worth having. A knowledge graph addresses all three. It holds a persona as a set of labelled facts, called triples, each in entity–relation–entity form: (persona, works_as, pharmacist). Knowledge graphs are valued for their symbolic reasoning ability and because new facts can be continuously added to them, as set out in a 2024 roadmap [5]; applied to a persona, a graph gives multi-step reasoning methods explicit links to follow, and a single fact can be added or changed without rewriting the profile. The design question, on this view, shifts from what the persona says to how its facts reach the model: written into its weights, or supplied to it at the moment of use.

3. Where an Identity Is Written

3.1 Feed-Forward Layers as Fact Memory

Nearly every method can be characterised by two questions; [5] surveys many of the techniques involved. First, how is the persona represented when it enters the model — flattened into ordinary tokens, encoded as dense vectors, passed through a graph network, or retrieved and placed in the prompt. Second, where in the transformer does that representation arrive: at the input embeddings, in the attention layers, in the feed-forward block, in the residual stream, or at the final output.

Answering the second question well rests on one of the field's most consequential findings. Feed-forward layers behave like key–value memories — the first weight matrix detects a pattern in the input, the second promotes the output tokens associated with it, a result published in 2021 [6]. Causal tracing in 2022 then found that, in GPT-style models, middle-layer feed-forward modules play a decisive role in mediating factual predictions, and showed their weights could be edited directly to plant a new fact [7], an approach later scaled up in 2023 to insert thousands at once [8]. For weight-based methods, that makes persona injection look more like writing to memory at a particular address, though these studies concern factual associations in general, not persona knowledge specifically. One study found that injecting knowledge into the feed-forward block outperformed adding it at the input [9].

3.2 The Graph-Based Camp

A graph-based camp sits right beside the weight-editing one. Hybrid architectures pair a language model with a graph neural network so that structure and text reason together, exchanging information through multiple dedicated fusion layers and outperforming models eight times larger on commonsense and medical question-answering benchmarks (CommonsenseQA, OpenbookQA and MedQA-USMLE), an approach published in 2022 [10]. Work in the same spirit fuses entity representations into the encoder itself, published in 2019 [11], or attaches triples to the input as small trees of tokens, published in 2020 [12]. Each answers the same "where" question with a different address.

4. Trained In, or Served Out

Here the two halves separate. A persona can be placed in the model before it runs, by changing something in the weights, or supplied while it runs, by changing what the model is given to work with.

4.1 Editing the Weights

On the training side, editing weights writes a fact into the parameters themselves, so it persists across sessions and adds no inference-time overhead, although later training or further edits can still disturb it; an approach introduced in 2022 and extended in 2023 [7], [8]. A complementary approach, published in 2022, leaves the base model frozen and trains a small low-rank adapter instead, so one shared base model can serve many small task-specific adapters, switched by swapping them in and out [13]. Applying that pattern to personas, which is our extension rather than a result of the original work, would give one model a roster of identities, though keeping them isolated from one another remains a design problem in its own right. A related adapter-based approach uses the model's internal states to decide when to add new knowledge, reducing forgetting of what it already knows, as shown in 2024 [14].

4.2 Retrieval Without Touching the Weights

On the inference side, a newer line works without touching the weights at all. It keeps a per-user knowledge graph outside the model, retrieving and optimising it from feedback, and improves personalisation without changing model parameters, published in 2024 [15]. A related method, also from 2024, encodes a user's interaction history into embeddings that the model attends to through cross-attention, rather than describing the user in text [16]. A graph-based retrieval method built on user-centric knowledge graphs, published in 2025, was designed for, and evaluated in, settings where a user's own history is sparse or missing, and outperformed prior personalisation methods there [17].

5. What the Arrangement Is Worth in Practice

None of this has stayed in the lab, and the appeal is not only technical. One model, properly arranged, can do the work of a whole shelf of separately built and separately maintained systems.

Where the Value Shows Up

• One system, many roles. We argue that several role-personas on one model could let roles that differ in vocabulary, responsibility and tone share a single deployment, rather than each being served by a separately built system.
• Continuity on the other side of the screen. We argue that persona stability is what makes an assistant feel like the same assistant tomorrow as today, and that its value grows as more people come to rely on it.
• Cost that grows more slowly than headcount. Interchangeable adapters, or graphs held outside the model, mean adding a role need not mean funding another full model, per the 2022 adapter work [13], though storage, retrieval, inference and upkeep still grow with each role added.
• Separation you can inspect. Keeping each persona in its own graph makes what each role knows explicit and inspectable, and makes revising a fact a matter of editing a record, per the 2024 retrieval work [15]. Enforcing the boundary between roles still needs access controls on top.
• Knowledge that stays put. We argue that the facts and decisions recorded in a persona graph become an asset that outlasts whoever first supplied them, and can be handed on rather than rebuilt.

The significance of this work extends well beyond language models themselves. As organisations increasingly rely on AI to support employees, customers and specialised workflows, they need systems that can adopt different responsibilities without becoming separate systems altogether. Training and serving multiple personas from a single model offers a way to provide specialised behaviour while sharing the same base model and infrastructure, with each persona's own knowledge kept separate. Instead of maintaining many independent models, organisations can adapt one foundation model to different roles, making AI systems easier to deploy, update and scale as new requirements emerge.

The pragmatic argument reinforces the technical one. Retrieval-based persona systems leave the base model unchanged, which sidesteps the cost of retraining the base model and keeps an existing deployment usable exactly as it stands, as reported in 2024 and 2025 [15], [17] — a property that may make them easier to deploy.

6. Below the Layer, and Off the Server

Two threads are redrawing the map again. The first drops below the layer, down to the circuit — a particular combination of attention heads and feed-forward sub-layers that together perform one function. Interpretability research published in 2024 identified knowledge circuits in GPT-2 and TinyLlama that work together to express particular facts [18]. A steering method proposed in a 2026 conference submission, not yet an established result, targets circuits it identifies for personal facts and combines adapters with steering-based activation injection there; evaluated on synthetic personal knowledge graphs, it reports updating a user's facts without degrading existing knowledge [19]. If such results hold up more broadly, the choice of where to write could rest less on intuition and more on measurement.

The second thread concerns where the data sits. A federated design for personalised recommendation, published in 2026, keeps each user's personal knowledge graph on their own device and shares only small low-rank adapter updates, letting a lightweight model complete a private graph without centralising anyone's data [20] — an approach with obvious appeal wherever privacy regulation or plain trust is at stake.

7. In Short

Step back and the through-line is hard to miss. In under a decade the options have multiplied: describing a persona in a few sentences, writing facts into particular weights and layers, probing the circuits that express them, and keeping each identity in a graph outside the model, drawn on as needed rather than absorbed whole. These approaches coexist rather than replace one another, and each offers a different answer to how one network can serve many stable identities at once.

The momentum is what makes this such a rewarding area to follow. Active work is refining how personas are kept distinct from one another, how a new role can be brought up to strength from a small amount of history, and how the division of labour between layers can be measured rather than assumed — while the boundary between what is trained in and what is served out keeps moving as both sides advance. The question running through this work, should an identity live in the model's weights or alongside it?, is proving to be one of the more quietly profound problems in machine learning, and the answers are arriving fast.

And the answer will matter far past the page it's studied on — it decides whether one assistant can safely stand in for many different roles across a business.