# Mini Me: A Private Personal Agent — Founding Brief

Date: 2026-06-16

## What This Is

A personal AI agent that's a real extension of you: it knows your preferences, habits, and patterns well enough to act on your behalf in the real world, and it stays private by architecture, not by promise.

Pitch:

> "Order me my usual Chipotle."

It happens because the agent actually knows you. Nobody else sees the underlying personal model: not the company building it, not an ad network, not another AI training run.

## Why Big Tech Can't Just Ship This

Big tech already has most of the raw material: calendar, email, purchase history, chat logs, location, browsing, subscriptions. What they structurally cannot offer is a credible guarantee that knowing you that well will not be used against you.

Their constraints:

- Business models depend on attention, ads, cloud lock-in, or data advantage.
- Memory is trapped inside their apps.
- You cannot take what ChatGPT, Gemini, Siri, Copilot, or Alexa knows about you and move it anywhere.
- Company-accessible memory can be mined, subpoenaed, trained on, leaked, or repurposed.

The differentiator is not "we also have a personal assistant."

The differentiator is:

> This one is architecturally incapable of being used against you, and you can verify that.

That is the claim almost nobody else can credibly make because almost nobody else's business model survives making it fully true.

## What It Actually Does

Mini Me builds and maintains a private model of you:

- standing preferences,
- routines,
- recurring requests,
- favorite places/products,
- calendar patterns,
- tone/voice,
- personal constraints,
- past choices.

Then it acts on real tasks using that model, so you stop re-explaining yourself.

It gets better because it remembers what you choose. It does not get better by mining you for ads, training someone else's model, or enriching a platform-owned profile.

## Flagship Use Case: Order My Usual

Food ordering is the cleanest first proof.

The agent knows:

- usual restaurant,
- location,
- item,
- customizations,
- dietary constraints,
- typical timing,
- payment preference,
- pickup/delivery preference.

User says:

> "Order my usual."

Mini Me confirms if needed, places the order, and records the outcome privately.

This should be deliberately narrow. No connector sprawl. No general-purpose agent theater. One task, real execution, reliable result.

## How It Is Put Together

Two layers:

### Personal Model

The Personal Model is the actual "you" data:

- preferences,
- routines,
- voice,
- recurring requests,
- corrections,
- history of accepted/rejected actions.

It is encrypted and controlled by the user.

Rules:

- Never used to train shared models.
- Never sold.
- Never repurposed outside user-requested tasks.
- Exportable by user.
- Deletable by user.
- Visible to user.

### Execution Layer

The Execution Layer takes an instruction like "order my usual," pulls only what it needs from the Personal Model, and gets the real-world action done.

For the first use case, this can be one direct ordering path. It does not need a grand abstraction. It needs to place the order correctly.

The trust boundary is the product:

> Chipotle gets an order. Chipotle does not get your whole life.

## Privacy As Architecture

Privacy claim must be concrete:

- Personal Model is encrypted.
- User controls keys.
- No shared-model training.
- No ad targeting.
- No broad company access.
- Each task receives minimum necessary context.
- User can inspect, export, and delete memories.
- Access grants are scoped and revocable.

"We respect your privacy" is not enough. Every company says that. Mini Me must point at architecture and show why the claim is true.

## Chain / Rights Layer

The user's life can live on-chain as rights and commitments, not plaintext.

On-chain:

- ownership commitments,
- encrypted data hashes,
- access grants,
- revocations,
- provenance receipts,
- data licensing terms,
- audit logs.

Off-chain encrypted:

- actual preferences,
- raw history,
- app data,
- messages,
- location,
- calendar details,
- purchases.

This lets no platform own the data while letting the user choose who gets what rights.

## Private Agent-To-Agent Future

Mini Me becomes the foundation for personal agents.

Example:

> "Schedule coffee with Sam when we are both available."

Your agent and Sam's agent negotiate without revealing full calendars. They exchange only minimal derived facts:

- available windows,
- constraints,
- final accepted slot.

Same model later applies to:

- travel planning,
- group purchases,
- intros,
- family logistics,
- work coordination,
- health accountability,
- shared subscriptions,
- data licensing with companies.

The user chooses what information to share with big apps and companies. Apps request scoped access; user grants, denies, expires, revokes, or monetizes.

## MVP

Build one real path:

1. Capture one standing preference explicitly.
2. Store it privately.
3. Build one execution path for one food order.
4. Natural interface: "order my usual."
5. Confirm when risk is non-trivial.
6. Place order correctly.
7. Record private outcome.

No marketplace. No token. No multi-agent network. No generalized connector system.

## After MVP

Expand one task type at a time:

- recurring purchase,
- draft message in user's voice,
- schedule simple meeting,
- reorder household item,
- prepare daily brief,
- file/save useful context into personal wiki.

Personal Model should grow from real usage, not preference-form bloat.

## Business Model

Sell directly to person as private digital twin.

Natural model:

- subscription,
- optional local/private cloud sync,
- premium execution integrations,
- later team/family plans,
- later data-licensing marketplace where user controls and monetizes access.

Do not use ads or data brokerage. That kills the trust claim.

## Risks

1. **Trust**
   - "Private by architecture" must be inspectable and defensible.

2. **Execution reliability**
   - Real-world tasks have real failure. Wrong food order is failure, not bad chat.

3. **Big tech imitation**
   - They will add preference memory. Differentiation must stay architectural: where data lives, who sees it, whose model trains on it.

4. **Overgeneralization**
   - Building a universal agent too early will make the first proof weak.

5. **Privacy theater**
   - If company operators can read Personal Model, the core claim fails.

## Recommended Next Step

Build the single end-to-end use case:

- one real standing preference,
- one private store,
- one ordering flow,
- one natural command,
- one reliable outcome.

Proof question:

> Does this feel meaningfully different from Siri, ChatGPT, Gemini, or Alexa because it knows you privately and acts correctly?