BLUF

A personal intelligence does not need consciousness to alter a life. It needs enough longitudinal context to predict what matters and enough permission to act. This essay maps the emerging stack, proposes an authority ladder, and asks how the human keeps an enforceable right to say no.

Research Exploration. This essay connects released products, research prototypes, and a proposed architecture. A unified lifelong personal intelligence does not yet exist, and vendor or prototype capabilities should not be read as evidence of reliable autonomy.

At a predawn kitchen table, a person controls streams of calendar, health, document, message, and purchase context through concentric permission boundaries.

The system becomes consequential when context crosses the permission boundary and turns into action. The stop control remains closest to the person. Illustration generated for this essay.

The kitchen at 12:30 a.m.

The laptop fan is warm against the wooden table. A watch has already recorded five hours of sleep. The calendar holds a two o’clock meeting. Three versions of the same deck sit open beside a cold cup of tea.

An assistant with access to only the prompt can offer generic advice. An assistant with permissioned access to sleep, calendar, files, screen activity, and past outcomes can say something narrower:

Stop at 12:30. Eat. Walk outside. I pulled the three documents that changed the last decision.

Useful.

Then it says:

I moved the low-priority meeting at eleven.

Still useful, perhaps.

Then:

I cancelled the important meeting because you needed sleep.

The model did not become conscious between those sentences. It crossed an authority boundary.

That boundary is the real discontinuity in personal AI.

The pieces are already on the table

A finished personal intelligence still sounds like science fiction when described all at once: it knows your work, health, finances, relationships, habits, and plans; remembers what you see and hear; travels between devices; speaks to other agents; and intervenes before you ask.

The finished system does not exist. Its components increasingly do.

  • Google’s Personal Intelligence can, when enabled, connect Gemini with Gmail, Photos, YouTube, and Search. Google describes reasoning across those sources and retrieving specific personal details, with per-app controls.
  • Apple’s Siri AI combines personal context, on-screen awareness, app actions, and camera-based Visual Intelligence on supported devices. Apple’s developer material shows how app entities and intents expose content and actions to Siri.
  • ChatGPT Health can, for eligible users, connect Apple Health and supported medical records. OpenAI says connected health information requires permission by default and receives additional privacy protections.
  • Project Astra explores multimodal memory, tool use, personalization, proactive responses, cross-device continuity, and prototype glasses. Google DeepMind still labels it a research prototype for limited trusted testers.
  • VisionClaw, a 2026 research prototype, joined Meta Ray-Ban glasses with an always-on agent that could interpret egocentric context and initiate tasks such as creating events from posters, making notes from documents, or controlling devices.

These systems are not interchangeable. A product beta, a developer platform, and a laboratory prototype have different evidence standards. None establishes that a unified personal model is accurate, safe, or desirable.

Together, they make the architecture visible.

Seven layers of personal intelligence

Strip away the product names and the stack has seven layers.

Layer What enters the system What can go wrong
Perception Camera, microphone, location, watch, phone, glasses, ambient sensors Bystander capture, false recognition, continuous surveillance
Digital context Email, calendar, documents, search, messages, purchases, work systems Context collapse, stale records, excessive collection
Personal world model People, projects, goals, constraints, recurring decisions False relationships become durable assumptions
Longitudinal memory What changed, persisted, disappeared, or returned Selective memory becomes a distorted biography
Prediction Likely next actions, risks, needs, and conflicts Correlation is treated as destiny
Action Draft, send, book, buy, move, cancel, negotiate, control A suggestion becomes an irreversible consequence
Policy Scope, limits, confirmation, delegation, revocation, audit Permission expands silently or cannot be withdrawn 
flowchart LR
    A["Physical perception"] --> C["Personal world model"]
    B["Digital context"] --> C
    C --> D["Longitudinal memory"]
    D --> E["Prediction"]
    E --> F["Proposed action"]
    G["Policy and authority"] --> F
    F --> H{"Human confirmation required?"}
    H -- "yes" --> I["Ask with evidence"]
    H -- "no, within scope" --> J["Execute and log"]
    I --> J
    J --> K["Observed outcome"]
    K --> D

The arrows should not imply that more data always produces better assistance. A personal system can become more useful by learning what not to retain, which source to distrust, and when a blank field is safer than an inference.

Memory is not a folder

At the kitchen table, a photograph, a health record, and an unanswered message occupy different applications. The useful signal may live in the relationship among them.

A personal model might notice that a project reappears after conversations with one collaborator; late meetings correlate with missed workouts; a recurring purchase follows travel; or a person described as important has gone unanswered for twelve days.

Research on Human Digital Twins approaches part of this problem through physiological, behavioral, and environmental sensing collected from mobile, wearable, and IoT devices. The term remains broad. A health-oriented twin is not the same as a general model of a person, and continuous data does not automatically yield a faithful representation.

The difficult object is not storage. It is a versioned model of change:

event
→ inferred relationship
→ confidence
→ source
→ expiration
→ correction history

Every inferred relationship should carry uncertainty and a path for correction. “You tend to avoid calls after six” is not a trait. It is a provisional claim derived from a specific window of observations.

Prediction must remain defeasible

A good personal system can estimate ordinary patterns:

  • meetings after six often displace exercise;
  • travel increases discretionary spending;
  • sleep below a personal threshold precedes weaker concentration;
  • a recurring deadline produces the same avoidable rush.

Those estimates can support reflection. They become dangerous when the system turns them into identity.

The phrase “your behavior suggests what you actually want” deserves suspicion. Behavior reflects preference, but also illness, coercion, poverty, fatigue, habit, discrimination, and missing options. A model that sees a pattern may be observing a constraint.

The person must be able to say:

Your prediction is accurate. I am choosing differently anyway.

That sentence is not an error case. It is a core system requirement.

Authority is a ladder, not a switch

At 12:31, “help me prepare” can mean six different things. A responsible interface makes the rung visible before the system climbs it.

Level System behavior Default control
0 — Observe Read only the explicitly connected source Visible access indicator and short retention
1 — Explain Identify patterns and cite supporting records User can inspect and correct the evidence
2 — Recommend Suggest an action without preparing it No external side effect
3 — Prepare Draft, stage, or reserve without committing Human approves the exact artifact
4 — Execute reversible Move a low-priority meeting or place a cancellable hold Narrow scope, time limit, undo window
5 — Commit consequential Send, purchase, sign, disclose, or cancel Fresh confirmation and step-up authentication
6 — Restrict the human Refuse a purchase or disable an option Normally prohibited outside explicit legal or safety authority

The highest rung is not “more intelligent.” It is more coercive.

A simple execution gate can be written as:

\[\operatorname{Execute}(a,t) = I_{\text{scope}} \cdot I_{\text{fresh}} \cdot I_{\text{budget}} \cdot I_{\text{risk}} \cdot I_{\text{revocable}} \cdot I_{\text{confirmed}}\]

Each indicator is either 0 or 1 for action $a$ at time $t$. Multiplication is deliberate: one failed condition blocks execution. A sophisticated confidence score should not compensate for missing authority.

The policy itself needs versioning. If the spending cap changed from $20 to $2,000, the audit record should show who changed it, when, why, and which delegated agents inherited the new ceiling.

Agent identity is not enough

An agent approaching a bank or hospital needs an identity. But identity answers only one question: which software principal is speaking?

A personal proxy must carry a narrower constitutional record:

  • whom it represents;
  • which data it may read;
  • which facts it may disclose;
  • which actions it may attempt;
  • budget and time limits;
  • whether it may delegate;
  • which approvals are still required;
  • when authority expires;
  • how revocation propagates;
  • what happened while the person was absent.

This is why the Agent Passport idea becomes more than authentication. Recent NIST work on software and AI-agent identity explicitly raises delegation, update, and revocation questions. Emerging Agent Passport and Delegation Receipt work proposes signed receipts in which each delegation step can only narrow authority.

The core invariant should be:

\[\mathcal{A}_{child} \subseteq \mathcal{A}_{parent} \subseteq \mathcal{A}_{human}\]

A sub-agent can receive less authority, never more. Revoking the human’s root delegation must stop the descendants, including queued or asynchronous work.

Wearables close the prompt gap

Phones gave computers a rich record of digital life. Glasses move the sensor toward the user’s line of sight.

Today, a person asks:

What did the presenter say about revenue?

An egocentric system can potentially resolve which presenter and which meeting because it shared part of the physical scene.

Today:

How do I repair this?

The future system may see the component in the person’s hand.

Today:

Add this restaurant to my list.

The future system may notice the person paused outside twice—but whether it should retain that observation is a policy question before it is a memory feature.

A recent survey of AI smart glasses identifies lifelong personalized memory, proactive intelligence, and trustworthy egocentric systems among the central research challenges. The phrase “trustworthy” carries physical weight here. A camera worn on one person’s face also sees everyone across the table.

Continuous sensing therefore needs bystander rules, visible capture signals, private-space suppression, local processing where possible, and deletion that removes derived embeddings and inferences—not only the original video.

The body becomes another context stream

A watch leaves a trail of heart rate, movement, workouts, and sleep estimates. An AI can summarize that trail, but interpretation remains bounded by sensor accuracy, missing data, and clinical limits.

The Personal Health Insights Agent study demonstrated an LLM-based agent answering open-ended questions over longitudinal wearable data. The authors explicitly limited the aim: the system was not designed to answer specialized or complex medical questions beyond the scope of wearable data.

That boundary belongs in the personal-intelligence stack. Health context can support questions and preparation. It should not silently transform a consumer agent into a clinician, insurer, employer, or gatekeeper.

The kitchen-table assistant may say, “Your sleep estimate is unusually low relative to your recent baseline.” It should not leap from that trace to a diagnosis or disable a person’s choices.

The integrated model is more sensitive than any source

We usually regulate and secure personal data in separate rooms: medical records, purchase history, location, messages, search, employment files. The combined longitudinal model can reveal more than the sum.

Not:

Chinmay searched for X.

But:

Chinmay searched for X after speaking to Y, while working on Z, following three short nights, abandoned the idea two weeks later, and returned to it after meeting A.

The relationships become the asset—and the risk.

I use Human API as a deliberately uncomfortable metaphor for this layer: a machine-readable approximation of a person’s context, preferences, commitments, constraints, and history. It is not the person. It should never acquire that moral status. It can still become useful enough that other systems try to query it.

The architectural choice determines where power sits:

  1. A platform holds the integrated model.
  2. Several providers maintain competing partial models.
  3. The individual holds an encrypted, portable model and issues temporary scoped access.

The third option is attractive but difficult. Portability requires common schemas. Local control requires usable key recovery. Derived models may leak even when raw records remain encrypted. A small device cannot carry every computation. “User-owned” becomes an empty phrase unless the user can inspect, export, revoke, and delete both data and inferences.

When personal agents negotiate

At an airline counter, a human compresses a complicated situation into a fare class and a few sentences. A personal agent could represent more variables: time flexibility, mobility needs, loyalty status, budget, cancellation tolerance, and who may see each preference.

Extend that pattern:

  • a candidate agent negotiates with a hiring agent;
  • a patient agent prepares questions for a hospital agent;
  • an employee agent coordinates with an organizational agent;
  • a purchasing agent compares sellers within a budget and disclosure policy.

This does not make money obsolete. Prices remain powerful summaries of scarcity. It suggests that agents may negotiate across more dimensions before a price or contract is settled.

The speculative branch is decentralized coordination:

human objectives + constraints
→ scoped personal agent
↔ institutional agents
→ proposed agreement
→ human or policy approval

The danger is equally clear. If platforms own the personal model, the negotiation may optimize the platform’s objective while appearing to represent the human. An agent that knows a buyer’s urgency can serve the buyer—or extract the maximum price.

Representation requires fiduciary-like clarity: whose interest is being optimized, which conflicts exist, and which information stays private during negotiation.

A research program for the permission boundary

The useful next step is not a demo with every connector enabled. It is a sequence of narrow trials.

Research or industry partner First artifact to test
Privacy or usable-security researcher A consent interface that explains source, inference, retention, and downstream disclosure
Identity and authorization team Delegation receipts, spending caps, expiry, and cascade revocation across sub-agents
Wearable or smart-glasses lab Bystander-aware capture and deletion of raw and derived data
Health-informatics group Bounded wearable reasoning with uncertainty and escalation to professional care
HCI or behavioral-science lab Whether users distinguish recommendation, preparation, and execution under time pressure
Personal-data or local-first infrastructure team A portable encrypted model with provider-independent export and verifiable deletion

One experiment could stage the same calendar conflict across the authority ladder. Participants see identical recommendations, but the system alternates among suggesting, drafting, moving with undo, and moving without confirmation. Measure comprehension, correction, regret, and whether users can accurately describe what the agent was allowed to do.

Another trial could test longitudinal inference. Show users the source events behind a prediction, let them correct the relationship, and return months later. Does the system preserve the correction, or does the same false inference reappear?

A narrow bridge to Aegis

This article is an independent personal-AI proposal. Its authority ladder also fits the governance work in Aegis: evaluate the model, but govern the action through scope, evidence, escalation, logging, and human override. Aegis would contribute enforcement patterns; it would not own the person’s life model.

The right to be inconsistent

Personal intelligence will probably arrive without a launch event.

First it knows the inbox. Then the calendar. Then files, health, screen context, and the object in view. Memory stretches from a conversation to a year. Small actions become routine. A travel hold becomes a purchase. A draft becomes a sent message.

No single step requires “Personal AGI.” The system becomes consequential through accumulation.

The final safeguard is not a smarter prediction. It is an enforceable human right to interrupt the model’s story:

I know what the evidence says. I understand the consequence. I am doing it anyway.

Sometimes that choice will be unwise. Human autonomy includes unwise choices within the boundaries of law and the rights of others.

The research problem therefore sits between agents, human digital twins, multimodal wearables, personal knowledge graphs, authorization, privacy, economics, and governance. All of those fields meet at one physical moment: a finger over the confirmation control while the system believes it already knows the answer.

How do we build an intelligence that can know a great deal about a person, operate across their digital world, and communicate with other agents—without quietly becoming the thing that runs their life?

The sensing is arriving. The memory is taking shape. The tools can act.

The unresolved design is permission.

References

After the last line

Comments & sharing

Agree, disagree, add context, or send this to someone who would have a take.

Share

LinkedIn X Email

Public thread

Comments are public and attached to this post through GitHub Discussions. Sign in with GitHub to join the thread.

Open public discussions

Newsletter, eventually

Get the next brain dump.

No fixed cadence yet. Leave your email for the first issue when it exists.

One list. No schedule. Unsubscribe whenever. How your email is handled.