XIOM

Introduction

The gap isn't in what AI can do.
It's in what AI is allowed to do.

AI models advance monthly. The way you use them hasn't changed in two years.
You ask. It answers. You forget. It forgets.
The most powerful technology in history, reduced to a glorified search bar.

Today's AI What's Possible

01

Waits for your prompt

→

Initiates when it matters Proactive delivery before you ask

02

Black box decisions

→

Fully explainable reasoning Every action has a receipt

03

Generates text

→

Executes real outcomes Actions, not drafts

04

Static responses

→

Continuous co-evolution Learns and improves with every interaction

05

Their platform, their rules

→

Your system, your data Local-first, portable, sovereign

06

Forgets everything

→

Remembers what matters Persistent world model that compounds

Why This Gap Exists

It's not a technology problem — today's AI models are extraordinary. The limitation is architectural: they're designed as products you visit, not partners who know you. They optimize for engagement metrics, not your life outcomes. You adapt to their interface — they never adapt to your reality.

XIOM changes what AI is allowed to do.

Not a chatbot. Not an app.
A persistent, private, proactive intelligence layer
that works for you — even when you're not looking.

1 — THE PROBLEM

Everyone has the same AI.
If your value is prompting — you're competing with everyone.

THE MATH

Why you have no moat

Your "AI advantage"

Same models as everyone 0 moat
Prompts anyone can learn 0 moat
Context in your head, not a system 0 moat

Your labor

Re-explaining every session replaceable
Copy-paste-verify loop replaceable
Manual follow-ups replaceable

Re-explain → Generate → Copy → Paste → Verify ↻ Repeat

∞ YOU

Every hour in this loop is an hour anyone else could do.
No compounding. No leverage. Just labor.

THE TAX

What this loop actually costs you

Time tax

+2h / week

You spend hours re-explaining, fixing, and following up. None of it compounds.

Attention tax

constant interrupts

You stay in the loop as executor, verifier, and reminder. Context switching becomes the job.

Failure tax

missed outcomes

Follow-ups slip. Deadlines drift. You only notice when it's already late.

THE TRAP

What you don't own

✗

Context

Your history lives in their UI. Switch vendors — lose everything.

✗

Rules

No enforceable policies. Every session is a new negotiation.

          if (amount > $500) → ask

          // ignored every session

✗

Loop

Drafts aren't results. You copy, paste, send, check, fix.

Draft v1 → Draft v2 → Draft v3 → ...

⚠ A junior with the same prompts produces the same output. A template could replace your "process."

If you are the glue — you're the part that gets replaced.

THE SHIFT

Own the infrastructure, not the chatbot

○ RENTING

Context locked in vendor UI
Rules reset every session
Manual copy-paste loop
No audit trail
Vendor lock-in

● XIOM

Memory portable, versioned, yours
Policy rules that persist
Execution — actions, not drafts
Audit receipts for everything
↻ Model swap anytime

Assistants are an interface. XIOM is infrastructure.
Interfaces talk to you. Infrastructure carries responsibility when you forget.

THE ESCAPE

From replaceable to irreplaceable

REPLACEABLE

You are the system

Memory lives in your head
Rules reset every session
Execution is manual labor
Anyone with prompts = same output

IRREPLACEABLE

You own the system

Memory compounds over time
Rules persist and enforce
AI does the labor
Your architecture is unique

2 — PRODUCT

The governance layer for personal AI.

Policy-controlled actions. Persistent context. Auditable outcomes.

○ TYPICAL ASSISTANT

... Waited for prompt

... Lost prior context

✗ No action executed

PROMPT-DRIVEN · SESSION-BASED CONTEXT
SUGGESTS TEXT, NOT OUTCOMES

● XIOM SYSTEM

✓ Checked calendar

✓ Proposed reschedule

✓ Requested approval

✓ Sent invite

✓ Logged receipt

EVENT-DRIVEN · PERSISTENT STATE AND MEMORY
EXECUTES WITH VERIFICATION

THE GOVERNANCE LOOP

Observe. Decide. Act. Verify. Learn.

Outcome governance for personal workflows. Policy-bound. Measurable. Reviewable.

Governing

Your Outcomes

24/7 · Autonomous

STEP 1 Observe — Calendar, health, memory, signals

STEP 2 Decide — Priority + authority check

STEP 3 Act — Execute or request approval

STEP 4 Verify — Confirm outcome achieved

STEP 5 Learn — Update policy + memory

Key difference: AI agents do tasks. XIOM governs outcomes.

SIX PROBLEMS SOLVED

What XIOM addresses.

01 INITIATIVE

"AI waits for prompts"

Goal-triggered actions (within policy)

02 TRANSPARENCY

"Black box decisions"

Intent + Context + Policy → Action (auditable)

03 EXECUTION

"Text suggestions only"

Outcome execution with verification

04 ADAPTATION

"Static configuration"

Versioned policies with feedback loops

05 CONTROL

"Platform lock-in"

Local-first, exportable, authority tiers

06 CONTINUITY

"Session amnesia"

Constitutional Memory with drift detection

AUDITABILITY

Every action is reviewable.

Action triggered: Goal drift detected → Priority alert sent

RECEIPT #2026-01-04-0847

INTENT Notify user of goal drift

CONTEXT Goal "Ship XIOM" · 3 blockers · Last session 17d ago

POLICY Alert if goal drift > 3 days v14

ACTION Priority alert sent ✓

APPROVAL User confirmed re-prioritization

ROLLBACK Available

AUTHORITY

Set permission tiers per workflow.

Observe Read-only access. Monitors signals, surfaces insights. Never acts without permission. Read Only

Suggest Proposes actions for your review. Executes nothing until you approve. Passive

Confirm Executes pre-approved action templates. Asks for one-time confirmation per new pattern. Gated

Supervised Runs autonomously within defined scope. Flags edge cases. Full audit log. Scoped

Autonomous Full policy-bound execution. Self-verifies, self-corrects, learns from outcomes. Full

THE GATE

Automation ships only when performance improves.

σ = P(h+XIOM) − max(P_h, P_AI)

σ > 0

Ship

Human + XIOM outperforms either alone

σ = 0

Keep Manual

No measurable improvement — no change

σ < 0

Block

Automation degrades outcomes — rejected

THE ARCHITECTURE

Four primitives that connect the system.

01 Constitutional Memory

Goals, constraints, and history — with full provenance. Versioned. Portable. Yours.

02 Governance Engine

Observe, decide, approve, act, verify, learn. The loop that never stops.

03 Execution Layer

Tool actions via scoped permissions and workflow policies. Actions, not drafts.

04 Receipts

Intent, context, policy, action, result. Exportable. Reviewable. Anchored on Base.

Ready to own your AI infrastructure?

Built for founders, operators, and creators with too much context.

Request Early Access →

3 — Operational Taxonomy

They're all building agents.
We're building the next level.

An operational taxonomy of AI autonomy — from informational systems to symbiotic AI. Five levels. Only one creates true partnership. XIOM operates at Level 5.

Level	Step Trigger	AI Initiative	Execute	Auto Loop	Persistent Memory	Policy Update	Audit	Scientific Class
L1 Informational AI	User	✗	✗	✗	✗	✗	✗	Non-agent information systems
L2 Reactive AI	User	✗	✗	✗	✗	✗	✗	Reactive agents, interactive assistants
L3 Mixed-Initiative AI	User + AI alerts	✓	✗	✗	◊	✗	✗	Mixed-initiative interaction (HCI + AI)
L4 Delegated / Agentic AI	AI within task	✓	✓	✓	◊	✗	◊	Goal-based agents under supervision
L5 Symbiotic AI	Long co-adaptive loop	✓	✓	✓	✓	✓ by outcomes	✓	Co-adaptive systems, human–AI teaming

L4 Agentic AI

LifecycleTask-bound → dies when task ends

InitiativeWithin the assigned task only

User Model◊ Optional, rarely maintained

Policy Update✗ None — resets per session

L5 Symbiotic AI

LifecyclePersistent → lives alongside you

InitiativeProactive, cross-context, goal-driven

User Model✓ Mandatory — the world model

Policy Update✓ By outcomes — self-improving

WHERE BIG TECH STANDS

They have the resources. They don't have the architecture.

L2 – L3

ChatGPT App

Reactive chat. Memory opt-in, siloed per account. No persistent policy enforcement.

L2 – L3

Claude App

Conversational assistant. Projects feature partial continuity. No execution layer.

L2 – L3

Gemini App

Google-integrated reactive AI. Cross-app data access but no user-governed policy layer.

L4

Operator (OpenAI)

Task-scoped browser automation. Executes within session. No persistent user model.

L4

Computer Use (Anthropic)

Desktop task execution. Powerful but stateless. No constitutional memory or auditing.

L4

Project Astra (Google)

Multimodal agentic assistant. Impressive perception, limited cross-session persistence.

STRUCTURAL BARRIERS

Why incumbents can't ship L5.

Stateless Architecture

Built for per-session inference. Persistent state requires full re-architecture.

Centralized Data

Deep user models require data they can't legally hold in most markets.

Policy Update = Liability

Self-modifying AI behavior creates regulatory and safety risk at scale.

Deep User Model = Privacy Risk

Storing behavioral models on central servers conflicts with GDPR and user trust.

Revenue Model Conflict

Subscription engagement metrics conflict with "do the task and step back" design.

Scale vs Depth

L5 requires per-user depth. At 100M users, depth is operationally impossible.

4 — The Science of L5

Symbiotic AI is not a product decision.
It's a scientific claim.

XIOM is grounded in six research traditions spanning cognitive science, control theory, and human–computer interaction. The L5 framework is falsifiable, measurable, and peer-reviewed.

"A Symbiotic AI system maintains a persistent, bidirectionally updated model of the user, takes proactive goal-aligned initiative across contexts, executes actions under policy governance, and adapts its own policies based on verified outcomes — in a continuous co-adaptive loop with a specific human partner."

— XIOM L5 Definition · Operational Taxonomy of AI Autonomy, 2026

THE FOUR AXIOMS

What makes an AI system truly L5.

A1

Mandatory User Model

The system must maintain a persistent, structured model of the specific user — including goals, constraints, preferences, and behavioral patterns. This model is not optional. Without it, the system cannot be L5.

A2

Bidirectional Policy Update

Policy governing AI behavior must be updated based on outcomes, not just instructions. The system learns what works. Rules evolve. The user's preferences are encoded structurally, not re-stated each session.

A3

Co-Adaptation Loop

Both the human and the AI change over time in response to each other. This is not personalization — it is genuine adaptation. The system's model of you diverges from any population average.

A4

Divided Subjectivity

The AI maintains distinct representation of its own state, the user's state, and the shared context. It can reason about the difference. Confusion between these leads to hallucinated alignment.

SYNERGY FORMULA

When does the system ship?

XIOM measures performance improvement before enabling autonomous action. If the combined system does not outperform either alone — it does not act.

S = P(h+ai) − max(P_h, P_ai)

P(h+ai)

Performance of human + XIOM system combined

P_h

Performance of human alone (baseline)

P_ai

Performance of AI alone (no human)

SYNC QUALITY METRIC

How well are human and AI aligned?

Sync (σ) measures the quality of human–AI coordination across four dimensions. All four must be present for L5 to function.

σ = (C · F · Q · K)^1/4

C — Context

Shared situational awareness

F — Feedback

Loop speed and accuracy

Q — Quality

Outcome verification fidelity

K — Knowledge

User model depth and currency

RESEARCH FOUNDATIONS

Six traditions that define the field.

A1 Cognitive Extension Clark & Chalmers (1998) — external systems as part of cognition. The world model is an extended mind, not a tool.

A2 Mixed-Initiative HCI Horvitz (1999) — systems that proactively take initiative based on user context without explicit prompting.

A3 Control Theory Closed-loop feedback with policy correction. The governance engine is a control system operating on life outcomes.

A4 Goal-Based Agents Russell & Norvig — goal-directed autonomous systems. XIOM extends this with persistent cross-session goal tracking.

A5 Human–AI Teaming McNeese et al. (2021) — co-adaptive joint cognitive systems where both agents update shared mental models.

A6 Synergistic AI Shneiderman (2020) — human-centered AI that enhances rather than replaces. XIOM's sigma gate operationalizes this.

5 — Memory Architecture

Your context is yours.
It lives where you decide.

Constitutional Memory is not a database. It's a structured representation of who you are, what you care about, and what your AI is allowed to do on your behalf.

THREE ARCHITECTURAL PILLARS

How XIOM stores what matters.

⟳

Persistent Memory

Neo4j Graph Database

Goals, constraints, relationships, and behavioral history stored as a versioned knowledge graph. Every node has provenance — who created it, when, and why.

◉

User Model

Constitutional Graph

A structured representation of your values, priorities, constraints, and preferences. The governance engine reads this before taking any action.

↺

Policy Update

Outcome-Driven Insights

Rules that work get reinforced. Rules that fail get flagged for review. Your constitution improves over time — not by instruction, but by outcomes.

OWNERSHIP PRINCIPLES

What "yours" actually means.

Portable

Your world model exports as structured JSON. Switch AI models, switch devices, switch everything — your context travels with you.

Versioned

Every change to your world model is versioned. Roll back to any prior state. See exactly when a rule was added and what triggered it.

Provenance

Every fact has a source. Every policy has an author. Nothing is added to your constitutional memory without a traceable origin.

What Is Stored

Goals and sub-goals with priority scores
Constraints and hard rules (never do X)
Preferences and soft preferences
Decision history with receipts
Entity relationships (people, projects, contexts)
Behavioral patterns and drift signals

What Is Not Stored

Raw message content from AI conversations
Biometric or health data without explicit opt-in
Third-party data without provenance
Anything you delete — immediately purged
Inferred data not confirmed by you

DATA CONTROL

You run the commands.

EXPORT PREVIEW

What your world model looks like.

// xiom-world-model-export.json
{
  "schema_version": "2.1.0",
  "exported_at":    "2026-07-29T12:00:00Z",
  "user_id":        "xiom_usr_a7f3c9",

  "facts": [
    { "id": "f001", "type": "goal",       "value": "Ship XIOM v1 by Q3",       "priority": 0.95 },
    { "id": "f002", "type": "constraint",  "value": "No decisions after 9pm",    "hard": true    },
    { "id": "f003", "type": "preference", "value": "Async over meetings",       "weight": 0.8  },
    { "id": "f004", "type": "pattern",    "value": "Deep work: 6am–10am daily", "confidence": 0.91 }
  ],

  "policies": [
    { "id": "p001", "rule": "Alert if goal drift > 3 days", "version": 14, "enabled": true },
    { "id": "p002", "rule": "Spend > $500 requires confirm",  "version": 7,  "enabled": true }
  ],

  "goals": [
    { "id": "g001", "title": "Ship XIOM",          "status": "active",   "progress": 0.72 },
    { "id": "g002", "title": "Raise pre-seed round", "status": "blocked",  "blockers": 3    }
  ]
}
    

6 — Roadmap

Phase 4 of 5.
The loop is closing.

Each phase operationalizes one step of the governance loop. We ship when the metric improves. We stop when it doesn't.

PHASE 01

Observe

✓ Done

What You Get

World model intake wizard
Goal + constraint capture
Daily signal monitoring
Constitutional Memory graph (Neo4j)

Delivered

Intake API + onboarding flow
Neo4j world model schema v1
User model node types
Signal ingestion pipeline

PHASE 02

Decide

✓ Done

What You Get

Priority scoring engine
Authority tier checks
Policy rule evaluation
Approval request flow

Delivered

Guardian pipeline v1
Constitutional rule DSL
Sigma gate (σ) metric
Approval notification system

PHASE 03

Act

✓ Done

What You Get

MCP tool execution layer
Calendar, email, file actions
Workflow automation engine
x402 payment protocol

Delivered

MCP server (JSON-RPC)
Tool permission scoping
USDG micropayment gateway
Rollback mechanism v1

PHASE 04

Verify

✓ Done

What You Get

Cryptographic receipts (Robinhood Chain)
Outcome confirmation loop
Audit log + export
AgentPassport NFT

Delivered

Receipt anchoring on Base
BidWall contract (revenue)
Full audit trail API
Dashboard receipt viewer

PHASE 05

Evolve

You are here

What You Get

→ Policy self-update from outcomes
→ World model drift detection
→ Co-adaptive learning loop
→ Full L5 Symbiotic AI runtime

Milestones

○ Outcome feedback pipeline
○ Policy versioning + diff engine
○ Drift alert + re-prioritization
○ Public beta launch

7 — Privacy

Your data is not our product.

XIOM is built on a simple premise: your AI should answer to you, not to our business model. Here is exactly what that means in practice.

OUR COMMITMENTS

What we guarantee.

[L]

Local-First

Your world model runs on your device by default. Nothing leaves your machine without explicit sync authorization.

[N]

No Monetization

We do not sell, share, or monetize your behavioral data. Our revenue comes from subscriptions and x402 protocol fees.

[P]

Full Provenance

Every fact in your world model has a source and timestamp. Nothing is inferred without your review.

[R]

Revocable

Delete any fact, any policy, any memory — immediately and permanently. No backup copies retained.

[A]

Audit Trail

Every action XIOM takes on your behalf is logged, timestamped, and available for your review indefinitely.

[S]

Sovereign Export

Your entire world model exports as structured JSON at any time. No proprietary lock-in. GDPR Article 20 compliant.

DATA WE PROCESS

What we store and why.

AccountEmail address, wallet address (optional), authentication tokens. Used for identity only.

ContentGoals, constraints, preferences, and facts you explicitly add to your world model.

Connected ServicesOAuth tokens for calendar, email integrations. Read-only by default. Revocable at any time.

System MetadataAction logs, receipt hashes, policy versions. Used for audit trail. Not used for analytics.

Working ModelThe Neo4j graph of your goals, entities, and relationships. Stored locally or in your private cloud.

TECHNICAL & LEGAL

The specifics.

Security

TLS 1.3 for all data in transit
AES-256 encryption at rest
Zero-knowledge receipt anchoring on Robinhood Chain
No plaintext storage of sensitive fields

AI Model Interactions

Prompts sent to AI providers are ephemeral
No training on your data without opt-in
Provider selection is yours — swap anytime
All AI calls logged in your audit trail

Data Retention

Account data: retained until deletion request
World model: yours — delete any node anytime
Audit logs: 90 days by default, configurable
Backups: 7-day rolling, then purged

Your Rights

Access: export full world model anytime (JSON)
Correction: edit any stored fact directly
Erasure: full account deletion within 24 hours
Portability: GDPR Article 20 compliant

Contact

Privacy questions: privacy@xiom-ai.com

Effective Date

July 29, 2026

Connect

Ready to own your AI?

XIOM is in active development. We're building the infrastructure for true human–AI partnership.

Founders Operators Creators Builders

Start in App →

"The future of AI is not assistance — it's partnership."

Download Desktop

Run XIOM on your machine.

XIOM Desktop hosts the MCP server locally, manages Neo4j, and pairs with your AI provider. Download a release build or run from source during development.

Windows

Installer for Windows x64. ARM64 build available below.

Download for Windows →

ARM64 · All releases

macOS

Disk image for Apple Silicon (.dmg).

Download for macOS →

All releases

Linux

AppImage not published yet. Check releases when available.

Coming soon →

All releases

Latest: desktop-v0.1.0

From source (developers)

git clone https://github.com/xiomAI-core/Xiom-ai.git
cd xiom
pnpm install
pnpm desktop

Requires Node 22+, Rust stable, and Tauri CLI. Desktop dev server runs at http://localhost:1420.

The gap isn't in what AI can do. It's in what AI is allowed to do.

Everyone has the same AI. If your value is prompting — you're competing with everyone.

Why you have no moat

What this loop actually costs you

What you don't own

Own the infrastructure, not the chatbot

From replaceable to irreplaceable

The governance layer for personal AI.

Observe. Decide. Act. Verify. Learn.

What XIOM addresses.

Every action is reviewable.

Set permission tiers per workflow.

Automation ships only when performance improves.

Four primitives that connect the system.

Ready to own your AI infrastructure?

They're all building agents.We're building the next level.

They have the resources. They don't have the architecture.

Why incumbents can't ship L5.

Symbiotic AI is not a product decision.It's a scientific claim.

What makes an AI system truly L5.

When does the system ship?

How well are human and AI aligned?

Six traditions that define the field.

Your context is yours.It lives where you decide.

How XIOM stores what matters.

What "yours" actually means.

You run the commands.

What your world model looks like.

Phase 4 of 5.The loop is closing.

Your data is not our product.

What we guarantee.

What we store and why.

The specifics.

Ready to own your AI?

Run XIOM on your machine.

Windows

macOS

Linux

The gap isn't in what AI can do.
It's in what AI is allowed to do.

Everyone has the same AI.
If your value is prompting — you're competing with everyone.

They're all building agents.
We're building the next level.

Symbiotic AI is not a product decision.
It's a scientific claim.

Your context is yours.
It lives where you decide.

Phase 4 of 5.
The loop is closing.