What a Personal AI Needs Instead of Just a Bigger Brain

Helm AI Blog · 2026-10-30 · The Person-Centered Series, entry 07

Devices change. Understanding should not have to start over.

The same person, restarting on every device

Picture one person across a single day.

In the morning, they use the AI on their phone to plan the schedule. At noon, smart glasses recognize a colleague walking past. In the evening, a home robot is asked to prep dinner.

Each device may have a camera, a microphone, a screen, actuators, location, and a general-purpose large language model. But none of them know:

• which colleague is allergic to shellfish;
• what the user's relationship with that colleague actually is;
• whether tonight's dinner is a casual meal or a formal hosted dinner;
• what the user already promised last week;
• which facts have quietly gone stale;
• whether the user even allowed the device to use this information.

So they remain general-purpose tools wearing different shells.

Two different problems: embodied vs. personal

The word "embodied AI" gets used for both warehouse robots and personal assistants. They are not the same job.

Embodied intelligence solves the physical world: perceive the environment, understand space and objects, predict physical change, plan actions, control a body, stay safe. A warehouse arm or inspection robot does not need to know your relationship history with a colleague.

Personal intelligence solves a different class of problem:

• Why is the user asking for this, now?
• Which people are involved?
• Who actually holds the decision?
• Who carries the consequence?
• What is the user's long-term aim?
• Does this request conflict with a past state?
• What fits this person, not the average person?

This distinction matters because it defines our relationship to embodied intelligence: we are not building the body or the physics engine. We are building the layer that tells the body who it is acting for.

"It can execute" is not "it understands"

Three requests, increasing in what they demand.

1. Pure execution: "Hand me the cup on the table." Vision, spatial reasoning, motion planning.
2. Personalized execution: "Make breakfast the way I usually have it." Now the system needs the user's habits, health limits, schedule, and recent changes.
3. Relational execution: "Have a few colleagues over this weekend, help me plan it." The system must know who counts as "colleague" here, who might come, each person's dietary restrictions, whether this is a thank-you dinner or a formal event, the role relationships, what it can use directly, and what it must confirm with the user.

A general model alone produces a "usually correct" banquet plan. A relationship world model is what makes a plan that fits this person and this group.

What a personal relationship world model actually contains

The name borrows from an established idea: a world model is a system's internal representation of the world it acts in. Embodied AI builds world models of physics and space; the paper we cite below, Embodied AI Agents: Modeling the World, uses exactly that term. We build a world model for the social and relational world a person lives in.

This is not invented from scratch. It draws on three recognized research lines — user modeling (how systems represent the people they serve), agent memory (how long-horizon agents store and retrieve context), and theory of mind (reasoning about others' beliefs, intentions, and relations). What stays under-served is the explicit, user-controlled treatment of relationships, authorization, and confidence. That gap is what a relationship world model is meant to fill.

It is not a social graph. At minimum it has seven parts:

1. The user model — long-term goals, current state, values, preferences and taboos, capabilities and limits, decision habits, authorized action boundaries.
2. People and relationships — who they are, what the user calls them, family/friend/colleague/client roles, relationship strength and change, situational roles.
3. Responsibility and power — who decides, who bears the consequence, who controls resources/rules/interpretation, whether the user can refuse, whether the relationship shifted.
4. History and time — when something happened, whether it still holds, whether a preference is a long habit or a temporary state, how views and commitments evolved.
5. Evidence and confidence — who the information came from (user statement, observed behavior, or system inference), whether the user confirmed it, whether conflicting evidence exists, current confidence.
6. Context and intent — what scene the user is in, the real task behind the surface request, which people/relations/history to invoke, and which memories should not be used this time.
7. Authorization and privacy — is this the user themselves, is it a temporary mode, what data may be called, what actions need re-verification, and how the user views, corrects, deletes, and withdraws authorization.

Memory is not enough

A plain memory system answers: what did we store?

A relationship world model also answers:

• who does this belong to;
• which person;
• in what relationship it holds;
• when it held;
• whether it still holds;
• what evidence supports it;
• what it has to do with the current task;
• whether the current device is allowed to use it;
• what to believe when evidence conflicts;
• whether the user must confirm.

The research is already converging

This is not a lone opinion. Four lines of work published in the last 18 months point the same direction.

• Embodied AI Agents: Modeling the World (Fung, Malik, et al., arXiv:2506.22355, 2025) argues that beyond the physical world, agents need a mental world model of the user for human-agent collaboration. The author list alone — Pascale Fung, Jitendra Malik, and others — signals this is now a mainstream research question, not a niche one.
• POLAR (Lee, Park, Lee; arXiv:2605.26256, 2026) organizes long-term interaction into a multimodal knowledge graph and shows the memory mechanism consistently improves performance, with the largest gains on multi-hop inference and tracking user-state change over time.
• Ella (Zhang, Gan, et al.; arXiv:2506.24019, 2025) uses a name-centric semantic memory plus spatiotemporal episodic memory to support long-term social interaction among 15 agents in a 3D world — relationships form and evolve, they do not come pre-coded.
• MEMENTO (Kwon et al.; ICLR 2026, arXiv:2505.16348) is the cautionary one. It shows that even frontier models like GPT-4o drop 30.5% in task success when they must reference multiple memories and user patterns — they fall back to generic "parametric commonsense" and ignore the specific user. Storing memory is not the same as using it correctly.

What we have actually built

We are not starting from a blank slide. Our memory layer already separates what most systems collapse:

• Source kind — is this the user's own words, observed behavior, or our inference? (This is the difference between a fact and a guess.)
• Truth status — a claim stays "unconfirmed" until the user confirms it; we do not silently upgrade an observation into a personality label.
• Authorization gates — only user- or externally-observed facts are explorable; model-generated content is quarantined from the user's real profile.
• Confidence portraits — every claim carries a confidence and a conflict trail.

What we are still researching — and are explicit about not having solved — is the top layer: a judgment layer that asks which tension in a person's situation is dominant right now, who controls the key resources and the interpretation of the rules — the authority and property relations a person is embedded in — and who pays the exit cost, and how all of that shifts over time. That layer is what turns "understanding" into "judgment." We are not claiming it is done.

Cross-device means a portable core, not a copied profile

The principle we are building toward:

• the user holds one encrypted, portable relationship core;
• each device receives only the minimum information the current task needs;
• every new observation enters as pending evidence, never directly as established fact;
• temporary modes do not leak into the permanent profile;
• important facts require user confirmation;
• the user can view, correct, delete, and withdraw authorization;
• multiple devices reconcile by identity, version, evidence, and time — not by overwriting.

Five risks we refuse to paper over

1. Inference as fact — one behavior is not a stable trait.
2. Correlation as authorization — knowing something is not permission to use it everywhere.
3. Personalization as control — the model should expand the user's choices, not make them.
4. Understanding as surveillance — a long-term model is only legitimate if it is authorized, visible, correctable, and deletable.
5. Personal data as platform asset — control of the relationship core belongs to the user, not to a device or a model vendor.

FAQ

Isn't this just a smarter memory system?

No. A memory system stores "what happened." A relationship world model explains "how what happened bears on who this person is and what they are trying to do now," and gates whether that history may affect the next action.

How is this different from embodied AI?

Embodied AI masters the physical world — perception, space, motion, safety. Personal intelligence masters the person — relationships, roles, history, intent, authorization. They stack; they are not the same layer.

Do you claim to have solved this?

No. The base layers (source separation, truth status, authorization, confidence) are built and in active development. The judgment layer is active research, not a shipped feature.

Why not just fine-tune a model on the user?

Fine-tuning bakes observations into weights where they cannot be inspected, corrected, or withdrawn. We keep understanding as explicit, auditable, user-controlled state instead.

What breaks if a device has no relationship model?

It produces "usually correct" outputs that are wrong for this person — the exact failure MEMENTO measured at 30.5% on frontier models.

Closing

Everyone may soon have an AI that stays with them for years. It will live first in a phone, then glasses, earbuds, home devices, robots. Its model will upgrade. Its hardware will change. Its body may migrate.

What makes it your AI is not how smart it is. It is whether it keeps understanding:

• who you are;
• what you have been through;
• who you are connected to;
• what you are actually trying to solve now;
• which decisions must stay yours.

Devices can be replaced. Understanding should not have to begin again.

References

• Fung, P., Bachrach, Y., Celikyilmaz, A., et al. Embodied AI Agents: Modeling the World. arXiv:2506.22355 (2025). arxiv.org/abs/2506.22355
• Lee, J., Park, C., Lee, D. Personalizing Embodied Multimodal LLM Agents over Long-term User Interactions (POLAR). arXiv:2605.26256 (2026). arxiv.org/abs/2605.26256
• Zhang, H., Zhang, Z., Wang, Z., et al. Ella: Embodied Social Agents with Lifelong Memory. arXiv:2506.24019 (2025). arxiv.org/abs/2506.24019
• Kwon, T., Choi, D., Kim, S., et al. Embodied Agents Meet Personalization: Exploring Memory Utilization for Personalized Assistance (MEMENTO). ICLR 2026. arXiv:2505.16348. arxiv.org/abs/2505.16348 · openreview.net/forum?id=vt4eO3NLE4

About the author

We started by trying to build a structured prediction engine — one that analyzes the central tensions a person faces in a given situation, and how those tensions shift over time. We hit a wall: you cannot predict a person's choices if you have not modeled the relationships those choices live inside. That pivot is why we now build a person-centered memory system first — get "understanding a person" right before prediction can stand on it. This article is the public version of that argument.

Read more about The Mirror →