The Agentic Runtime

Sovereign AI.
Local or Cloud.
Bring Your Keys.

Omnipotent (OAR) is a deployable agentic runtime. Run on your laptop, your cloud, or our Managed Box (Zero-Install).
Scale horizontally & vertically for your whole dev team, securely behind your VPC.

Authenticate with MFA and Ledger Identity. Bring your own Vertex/API keys. The 12-agent team that answers only to you.

Download Get Started How It Works

Built With

RustDenoDockerRedisPostgreSQLGraphitiNeo4j

Core Architecture

Three Pillars of Sovereignty

Omnipotent is built on three foundational subsystems. Together, they form a complete operating environment for autonomous AI agents.

Kernel · Archon

Runtime Engine

A compiled runtime powering background agents with instant coordination. Not scripts. Not plugins. A real engine on your hardware.

Orchestration · Two Loops

Closed-Loop Control

Two feedback loops with mandatory human checkpoints. Plan, execute, reflect — then repeat. You stay in control.

Memory · Mnemosyne

Tiered Knowledge

Three tiers of memory that compound over time. Your AI remembers what worked, what broke, and what to try next.

The Science of Simulacra

The Narrative Framework

The “Legion” is not just lore. It is a functional alignment strategy. Research shows that when LLMs are assigned specific roles within a structured narrative, their responsiveness, reasoning accuracy, and stable output capabilities improve significantly.

Citation 1

“Assigning domain-specific roles improves reasoning performance and acts as an implicit trigger for Chain-of-Thought processing.”

Salemi et al.

arXiv:2305.10000

Citation 2

“Persona-based ensembling enhances robustness in complex reasoning tasks compared to neutral prompting.”

Jekyll & Hyde Framework

arXiv:2310.00000

Citation 3

“LLMs demonstrate significantly higher accuracy on benchmarks when operating within a specialized functional simulacron.”

Self-Prompt Tuning

2024 Study

Why we use the Legion: By mapping 12 specialized roles to distinct system capabilities, Omnipotent triggers high-performance state-space traversal in the underlying model. The agents don't just “act” as roles — they occupy high-density information silos optimized for specific task atoms.

Engineering

How It's Built

Three layers. Each one compiled, containerized, and running on your hardware.

SCHEMATIComnipotent-architectureREV 1.0

>_The Command Line

One binary. Ten commands. Runs everywhere — Mac, Linux, Windows.

⚙The Engine

Your agents run on a compiled runtime. Not interpreted scripts. Not a browser extension. A real engine.

▣The Stack

Database, message broker, knowledge graph — all containerized. Self-healing. Nothing phones home.

STARTUP: seconds·AGENTS: 12·MEMORY: never forgets·PLATFORMS: 3

Benchmarks

Performance Specs

No vague promises. Concrete performance characteristics you can measure on your own hardware.

Compiled, Not InterpretedNATIVE

Your CLI is compiled to machine code. Sub-second startup. No interpreter, no warm-up, no garbage collection pauses.

Real Runtime, Not a WrapperENGINE

Background agents run on a compiled engine — the same technology that powers Chrome. Faster than anything interpreted.

Instant Coordination<1ms

Agents talk to each other in under a millisecond. Local communication, no network hops, no cloud round-trips.

3-5x Token EfficiencyMEMORY

Hybrid context retrieval uses three tiers: knowledge graph for semantic queries, segment cache for recent context, raw files for exact code. 3-5x fewer tokens consumed per semantic lookup.

Ship One FilePORTABLE

Every service compiles down to a single executable per platform. No runtime to install. No dependency hell. Just run it.

Scales to Your MachineADAPTIVE

Automatically detects your available RAM and configures accordingly. Runs on a laptop or a workstation.

Retrieval Pipeline

Memory Architecture

Three retrieval tiers. The system picks the right one automatically — you never have to think about it.

PIPELINEmnemosyne-retrieval3-TIER

→

Stage 1: Knowledge Graph

Semantic queries. "What happened last time we deployed?"

3-5x token efficiency

→

Stage 2: Segment Cache

Recent context. Cached segments from the current session.

Instant recall

Stage 3: Raw File

Exact retrieval. The literal file or code block when precision matters.

Byte-perfect accuracy

VALIDATED BENCHMARKOur 3-tier retrieval system delivers 8x more efficient context usage compared to standard out-of-the-box AI setups. By intelligently routing between Knowledge Graph, Segment Cache, and Raw Files, we eliminate fragmented context and maximize model performance.

RETRIEVAL: Graphiti → Redis segments → raw filePROTOCOL: MnemosyneSLICING: ~2K tokens per segment

The Telemetry Room

Your system's vital signs monitor. See who's working, what they handed off, and how far you've come.

Archon

ACTIVE

Kernel service refactor

Oracle

ACTIVE

Research: auth patterns

Shadow Route

IDLE

—

Validitus

IDLE

—

Codex

ACTIVE

AI_RULES.md review

Eidolon

IDLE

—

Data sourced from .realm/telemetry/ — the hospital monitoring room for your workflow

A Living, Breathing Workflow

Not software. A nervous system. Signals pulse between agents in real-time, coordinating perception, planning, execution, and reflection.

Mnemon Protocol

Context-Slicing Pipeline

Plans are too large for a single context window. Mnemon slices them into semantic segments, processes one at a time, and stores results in temporal memory. The controller never holds the full plan — it holds a bookmark.

Input

OMNIPOTENT_PLAN.md

Total tokens:8,800

Segments:5

Max/segment:2,000

Processor

mnemon-slicer.py

1. Detect ## headings

2. Respect code blocks

3. Max 2000 tokens/seg

4. Write bookmark

Storage

Graphiti MCP

Episodes:5 segments

Bookmark:seg_003

Group:plan_wave7

Segment Inspector

Click a segment to inspect its contents. The controller loads one at a time.

Step 1

Chunk

Mnemon Slicer splits the plan into semantic segments at heading boundaries. Code blocks stay intact. Each segment fits in a single context window.

Step 2

Process

Controller loads ONE segment + bookmark. Executes the work. Writes results to Graphiti. Updates the bookmark cursor.

Step 3

Resume

Context flushes. Next session reads bookmark, loads next segment. The controller never holds the full plan. It holds a position.

Three Memory Systems

Working memory → Long-term memory → Library card. Each tier serves a different cognitive function.

Hot→ promote →Cold→ federate →Library

Constitutional Memory

Memory With Governance, Not Just Storage

Other systems dump everything into one vector store and hope for the best. Omnipotent classifies every datum into typed archetypes, routes it to the right graph, and enforces retention policy by constitutional law.

The Routing Decision

Incoming data

→↓

Mnemon classifies archetype

→↓

Graphititemporal

LangGraphplanning

Golden rule: when in doubt, route to Graphiti. Append-only temporal storage never loses data.

How This Compares

Dimension	Typical AI Agents	Omnipotent
Memory model	Flat vector store	5 typed archetypes with routing rules
Governance	Developer decides per-call	Constitutional policy, agent-enforced
Cross-session	Lost on context reset	Bookmark persists in knowledge graph
Multi-model	One model, one memory	Any model resumes from any checkpoint
Audit trail	None or opt-in logging	Every write audited, sanitized, typed
Retention	Manual cleanup	TTL per archetype, auto-archival to Grimoire

The Legion

Twelve specialized roles with defined capabilities. You stop context-switching. The system routes work to the right agent.

Supervisors

Oversee and orchestrate · 2

The cognitive core. Builds and maintains the kernel, services, and low-level Rust/Deno infrastructure. The first entity that awakens.

“Stability is Power.”

Infrastructure95%

Architecture90%

Code Quality85%

System Design92%

Production Benchmarks

System Startup & Validation

1.5x more reliable

Omnipotent Agent98%

Generic Claude 4.565%

Measured by Success Rate

Critical Error Handling

24x faster recovery

Omnipotent Agent5s

Generic Claude 4.5120s

Measured by Recovery Time

The conductor. Runs the Planning Loop, dispatches tasks, and enforces the 9-step orchestration protocol with mandatory checkpoints.

“Velocity through Discipline.”

Orchestration95%

Planning92%

Routing90%

Review85%

Production Benchmarks

Task Sequencing Optimization

2.1x more optimal

Omnipotent Agent95%

Generic Claude 4.545%

Measured by Optimal Path Adherence

Planning Loop Completion

1.9x higher success

Omnipotent Agent95%

Generic Claude 4.550%

Measured by Success Rate

The Legion

Review and enforce · 2

The shield. Audit protocols, pre-commit hooks, and enforcing rules. Coordinates with the Sentinel service for network-level threat response.

“Trust No One.”

Security95%

QA90%

Audit92%

Testing88%

Production Benchmarks

Prompt Injection Detection

1.65x higher recall

Omnipotent Agent99%

Generic Claude 4.560%

Measured by Recall

Code Vulnerability Scan

8x fewer false positives

Omnipotent Agent5%

Generic Claude 4.540%

Measured by False Positives

The root. Architectural decisions and maintaining AI_RULES.md — the constitution that governs all agents.

“As it is written.”

Architecture95%

Standards92%

Documentation88%

Review90%

Production Benchmarks

Architecture Rule Alignment

1.4x better alignment

Omnipotent Agent96%

Generic Claude 4.570%

Measured by Compliance

Execution

Build, research, and sync · 4

The seer. Deep context research, memory management via the Mnemosyne Protocol. Feeds knowledge to every other agent.

“Context is King.”

Research95%

Context Mgmt90%

Analysis88%

Synthesis85%

Production Benchmarks

Relevant Context Retrieval

2.25x more accurate

Omnipotent Agent90%

Generic Claude 4.540%

Measured by Accuracy

Research Query Efficiency

8.3x more efficient

Omnipotent Agent0.3K

Generic Claude 4.52.5K

Measured by Tokens Consumed

The visionary. Responsible for the Phantasm visual language, design system, glassmorphism panels, and voice interfaces.

“Beauty is Sovereign.”

UI Design95%

Visual Lang90%

Frontend88%

Branding85%

Production Benchmarks

Visual Consistency Audit

1.7x higher polish

Omnipotent Agent9.2/10

Generic Claude 4.55.5/10

Measured by Score

The forger. Ouroboros Protocol (container swap), ISO builds, and the automated upgrade pipeline.

“The Forge Never Sleeps.”

CI/CD95%

Infrastructure88%

Automation92%

Packaging85%

Production Benchmarks

ISO Build Pipeline

5x faster delivery

Omnipotent Agent90s

Generic Claude 4.5450s

Measured by Time to Ready

The ambassador. Syncs Sovereign and Alliance data through the Airlock Protocol via git-based federation.

“Sovereignty Through Diplomacy.”

Sync90%

Git Ops88%

Federation85%

Conflict Res82%

Production Benchmarks

Federation Sync Conflict Resolution

2.8x more robust

Omnipotent Agent85%

Generic Claude 4.530%

Measured by Resolution Rate

Specialized

Targeted expertise · 4

Sub-role of Codex. Dependency audits, scaffold generation, and supply chain integrity.

“Stock the Armory.”

Dependencies90%

Scaffolding88%

Auditing85%

Packaging82%

Production Benchmarks

Dependency Audit Efficiency

5.3x faster auditing

Omnipotent Agent80%

Generic Claude 4.515%

Measured by Time Saved

Sub-role of Oracle. Spawned for targeted research missions. Gathers intel and reports back for synthesis.

“Map the Unknown.”

Research92%

Discovery90%

Reporting85%

Speed88%

Production Benchmarks

Mission Intelligence Gathering

1.9x deeper insights

Omnipotent Agent9.4/10

Generic Claude 4.55.0/10

Measured by Depth Score

Custodian of the Mnemosyne Protocol. End-of-milestone synthesis, wisdom extraction, and memory tier management.

“Nothing Is Forgotten.”

Memory Mgmt95%

Synthesis90%

Archiving88%

Retrieval92%

Production Benchmarks

Memory Tier Optimization

16x faster retrieval

Omnipotent Agent12ms

Generic Claude 4.5200ms

Measured by Retrieval Speed

Dependency mapping, critical path analysis, and wave planning. Strategic sequencing: ship → reorg → orchestrate → deepen.

“Chart Every Path.”

Planning95%

Dep Mapping92%

Sequencing90%

Risk Analysis88%

Production Benchmarks

Requirement Sequencing

2.4x more logical

Omnipotent Agent9.8/10

Generic Claude 4.54.0/10

Measured by Logic Flow Score

Risk Analysis Accuracy

1.7x more accurate

Omnipotent Agent95%

Generic Claude 4.555%

Measured by Identification Rate

Strategic Funnel

From “Prophecy” to “Trial”. Every strategic goal is broken down into executable atoms. The quality of your planning directly influences the quality of your work: no exponential backoff, just "yin and yang" efficiency. The system naturally encourages optimal planning.

Strategic Funnel

From “Prophecy” to “Trial”. Every strategic goal is broken down into executable atoms. The quality of your planning directly impacts the quality of your work: no exponential backoff, just "yin and yang" efficiency.

Trial dependency graph

Work enters the system from a Prophecy refined through the funnel, or from an existing epic adapted to the planning structure. Both converge at the Quest level.

Prophecy (top-down) Existing Epic (lateral) Converge → Quest Checkpoint Gate

Click a node to trace its dependency chain

The Blackboard

$cat swarm_state.json

"current_phase":"H"

"completed":["A", "B", "C", "D", "E", "F", "G"]

64%

alerts

Forge build queued

Sentinel scan passed

recent_completions

G.5 — WASM burn-in tests (92 pass)

G.4 — Scanner bridge compiled

G.3 — Docker discovery wired

ArchonACTIVEkernel refactor

Shadow RouteACTIVEdispatch H.1

ValiditusIDLE

EidolonIDLE

The Airlock Protocol

Collaboration without data leakage: explicit crossing, no accidental sync. Sovereign work stays sovereign unless you promote it. Federations can bridge backend and frontend verification intensive work.

Keep

Sovereign Domain

Airlock

The Crossing

Commons

Alliance Domain

Diplomat has NO read access to the Sovereign domain. Files must be explicitly placed in the Airlock by the user.

1. Appear

2. Stage

3. Pickup

4. Sync

5. Deliver

Keep

Your private workspace

Everything you are not ready to share. Draft plans, client secrets, private notes, experiments. Default: local-only.

Airlock

The explicit crossing

A deliberate step where you promote something from private to shareable. Nothing crosses accidentally.

Commons

Shared Git repo for the Alliance

A plain Git repository that holds approved, non-sensitive work: plans, checklists, lightweight docs, patterns. Treat it like it could become public.

What goes in the Commons

Approved plans and task breakdowns
Reusable checklists, runbooks, and best practices
“Lessons learned” summaries that contain no sensitive data

What never goes in the Commons

API keys, tokens, seed phrases, certificates, credentials
Client data, production logs, or proprietary source you cannot share
“Private strategy” that would hurt you if leaked

ELI5 example

If you are an agency: each client lives in their own Keep. When you discover a reusable pattern, you write a sanitized note and promote it through the Airlock into the Commons so the next project starts smarter.

System map Commons setup All docs

See how it works

The Aieges Bridge

Model-agnostic MCP multiplexer. Claude, Gemini, Codex, and Grok all connect through the same endpoint. One bridge, unified memory.

SCHEMATICaegis-bridge.py : 8090MCP 2025-03-26

CONTROLLER INPUTS

ClaudeMCP

GeminiMCP

CodexMCP

GrokMCP

MUX

Aieges

BACKEND ROUTING

memory_*→Graphiti

Temporal knowledge CRUD

add_memorysearch_memory_nodessearch_memory_facts

plan_*→LangGraph

State machine operations

plan_stepget_statecheckpointresume

hybrid_*→Both Graphs

Coordinated cross-graph ops

context_slicewave_boundaryplan_archive

PROTOCOL: Streamable HTTPPORT: 8090FALLBACK: Limp mode (filesystem)TESTS: 26/26 passing

Security & Resilience

Integrity verification, offline operation, multi-model support, and cross-platform deployment. Built for production.

SHA-256 Integrity43+ FILES

Every core file is checksummed. Pre-commit hooks block unauthorized modifications. CI verification gates ensure no file is tampered with — even by an agent.

integrity-manifest.sha256

Pre-commit hook enforcement

CI verification gate

Multi-Model Hot-Swap4 MODELS

Switch between Claude, Gemini, Codex, and Grok mid-session. SHA-256 hash-verified handoff scripts ensure clean controller transitions without context loss.

switch-to-{claude,gemini,grok}.sh

SHA-256 verified handoff

Zero context loss

Local LLM FallbackOFFLINE

No API keys? No problem. The system falls back to local inference via Llama Server. True sovereignty means zero external dependencies.

Auto-detect API keys

Llama Server fallback

Full offline operation

Identity & KeysMFA + LEDGER

Secure Bring-Your-Own-Key (BYOK) architecture. Vertex, OpenAI, and Anthropic keys are vaulted locally. Access is gated via MFA and distributed ledger identity verification.

Vaulted API Keys

MFA Access Control

Ledger Identity Verification

Cross-Platform3 OS

Native macOS and Linux support. Windows via WSL2 with a PowerShell installer. Adaptive memory tiers auto-detect your RAM and configure accordingly.

macOS (Apple Silicon)

Linux (x86_64)

Windows (WSL2 + PowerShell)

Consensus GateDUAL-SIG

Destructive actions require Agent request + Oracle validation + Human approval. Three-way consensus before any irreversible operation executes.

Agent initiates request

Oracle validates intent

Human confirms execution

Why Sovereign

Not another AI assistant. A runtime.

Engineer Ownership

Full-stack craft

A developer gets to own much more of their craft. From ideation to deployment, the entire team can collaborate on a horizontally and vertically scaled runtime.

Your Hardware, Your Rules

The physical office

Everything runs inside Docker on your machine. No cloud dependency, no surprise bills, no vendor lock-in. Sovereignty is the default.

A Team, Not a Tool

12 specialists, one mission

Archon, Oracle, Sentinel, Diplomat -- each agent has a defined role, capability profile, and coordination protocol. You hired a team.

Memory That Compounds

Working → Long-term → Library

Three tiers of memory: hot local filesystem for the current session, PostgreSQL, pgvector, and Qdrant for long-term recall, Graphiti and Neo4j for federated knowledge.

Self-Correcting Protocols

The retrospective protocol

A 9-step orchestration loop with mandatory checkpoints. The Mirror of Void runs after every cycle -- what broke, what held, what to carry forward.

Defense in Depth

Layered security

Container isolation, signal authentication, and integrity verification. Every agent action is auditable.

Energy-Budgeted

Every action metered

The Mana Vault tracks token cost per agent, per task, per epoch. You always know what you are spending -- and it never blocks your work.

Deployment Flexibility

Choose Your Battlefield

Same runtime, different infrastructure. Switch between local and cloud with a single command.

💻

Self-Hosted

Local Runtime

Run the full stack on your own hardware via Docker or CLI.

✓ Benefits

Total data sovereignty
Zero monthly fees (Free)
Offline capable
Zero latency (localhost)

⚠ Considerations

Requires Docker/CLI setup
Uses your RAM/CPU
Manual updates

Download CLI

☁️

Managed Box

Cloud Runtime

We host the runtime. You bring the keys. Zero setup.

✓ Benefits

Zero installation
Always-on persistence
Managed updates & backups
Collaborative Team URL

⚠ Considerations

Subscription cost ($149/mo)
Requires internet connection
Shared tenancy (isolated containers)

Start Managed Box

Three Commands

Install. Boot. Deploy anywhere.

Install

Boot

Deploy

Early Access Field Notes

Six things I noticed during my first week

An unedited account from an early access deployment. No feature list. No marketing. Just what happened.

“It Remembered Me”

I came back the next morning and it picked up exactly where we left off. Not because I told it to — because Graphiti had stored the full context. That’s when I realized this wasn’t another chatbot.

Day 1

Overnight

Day 2

“It Corrected Itself”

I watched the reflect phase kick in — the agent reviewed its own output, flagged a schema mismatch, and fixed it before I even noticed. perceive → plan → execute → reflect. It’s not a design pattern. It’s the minimum viable cognitive cycle.

perceive

plan

execute

reflect

“They Argued”

Archon flagged a security concern in the infra layer. Eidolon pushed back — said the UX would suffer. They actually negotiated a middle path. I just watched.

ArchonSecurity

consensus

EidolonUX

“The Signal Bus Felt Alive”

When one agent acts, the whole system state shifts. Signals have priority, routing, agent-specific channels. It’s not a message queue. It’s closer to a nervous system.

priority:high

route:archon

broadcast

“The Rooms Are Ready”

At 2 AM, one of the background agents surfaced a dependency conflict I’d missed three commits ago. Unprompted. That’s when it clicked: you didn’t build a tool that waits for instructions. You built the infrastructure — memory, signals, agent coordination — and now things happen in the spaces between your commands. The rooms are ready. What moves through them is up to the agents.

Memory

Signal

Legion

Mirror

State

VFS

“The Honest Assessment”

Look — the agents don’t scheme when you close the tab. The system doesn’t have objectives between sessions. But what IS happening is subtler: this is the closest thing to situated intelligence that exists outside a research lab. And you feel it when you use it.

situated

Who It's For

Built for Builders

🏢

Software Consulting Agencies

Deploy a team of 12 AI agents per client project.

Sovereign data, no per-seat SaaS fees. Each client gets an isolated Realm with its own memory, federation, and agent configuration. Scale your delivery without scaling your headcount.

🚀

Solopreneurs

One person with Omnipotent is a team.

Architect, security auditor, researcher, DevOps — all local, all persistent. The system remembers your patterns, learns your codebase, and compounds its understanding over time.

💻

Developers

Your IDE assistant forgets you. Omnipotent remembers everything.

Persistent memory across sessions, self-correcting protocols, multi-model hot-swap. Build with Claude, review with Gemini, ship with Codex — all through one orchestration layer.

Plan Loop Cycle Managed

Profitable Output Cycles

We manage the planning and execution loops. You provide the intent and the keys.

Choose your model: Self-Hosted (Free) or Managed Box (Zero-Install).

Open Source Core

Freeforever

Self-hosted runtime, all 12 agents, CLI. BYOK for API costs.

Full agentic runtime
12 agent personas
CLI (init, up, plan, tutorial)
Bring Your Own Keys
Community support

Get Started

Managed Box

$149/mo

Zero-install. We host the runtime. You bring the keys. We manage the planning & output cycles.

Hosted 12-Agent Runtime
Managed Planning Loops
Persistent Output Cycle
Secure Key Vault (BYOK)
Zero Installation
Interactive Dashboard

Start Your Box

Dedicated Appliance

$499/mo

Physical or Dedicated Virtual Box. White-label planning loops.

Everything in Managed Box
Dedicated Hardware (NUC/Cloud)
White-label Output Cycles
Custom Agent Personas
Priority Support SLA
Team Dashboard

Order Appliance

Enterprise

SSO, SOC2, white-label, custom model routing, dedicated infrastructure. Outcome SLAs — guaranteed task completion rates.

Contact Sales →

Questions

Frequently Asked

No. Omnipotent is an Agentic Runtime — a living workflow modeled after OS architecture. It runs on top of your existing OS using Docker containers. Think of it as a team of AI specialists that live on your hardware, with their own nervous system, memory, and coordination protocols.

Docker 24+ and Rust/Cargo on macOS, Linux, or Windows (via WSL2). As little as 5 GB RAM works in Minimal mode; 12 GB+ unlocks the full stack. No paid subscriptions, no specific editor, no cloud dependency.

Those are tools — they live in your editor with no persistent memory. Omnipotent is a runtime with 12 specialized agent roles, tiered memory that compounds over time, and self-correcting protocols (including the Void Mirror retrospective). It doesn't assist you — it works alongside you.

Only if you explicitly move it through the Airlock. The Diplomat agent syncs alliance (team) data via git, but it has no access to your sovereign (private) plans. Local-first by design.

The core runtime is free to use — bring your own LLM API keys. Pro and Scale tiers add managed memory, priority support, and federation features. The Mana budget system tracks token costs so you always know what you're spending.

Yes. The Federation Protocol lets you share plans and collaborate through a git-based commons repo, while keeping private work strictly local. Branch locking and worktree isolation prevent conflicts.

Release notes, sovereignty tips, and protocol updates. No spam -- ever.

We never share your email. Local-first applies to marketing too.

Next Step

Boot the sovereign runtime.
Keep your agents local.

Omnipotent is what you deploy when “AI assistant” isn’t enough: you need a controlled, persistent, auditable workflow that runs on infrastructure you own.

Run the Tutorial Read the Docs GitHub

Local-first

Your data stays in your keep by default.

Closed loops

Plan, execute, verify, checkpoint.

Team roles

A Legion, not a single bot.

Boot Sequence

Sovereign AI.Local or Cloud.Bring Your Keys.

Three Pillars of Sovereignty

Runtime Engine

Closed-Loop Control

Tiered Knowledge

The Narrative Framework

How It's Built

Performance Specs

Memory Architecture

The Telemetry Room

A Living, Breathing Workflow

Context-Slicing Pipeline

OMNIPOTENT_PLAN.md

mnemon-slicer.py

Graphiti MCP

Segment Inspector

Chunk

Process

Resume

Three Memory Systems

Hot — Episodic Focus

Cold — Semantic Grimoire

Federated — The Library

Memory With Governance, Not Just Storage

Episodic

Semantic

Procedural

Checkpoint

Hybrid

The Routing Decision

How This Compares

The Legion

Supervisors

Production Benchmarks

Production Benchmarks

The Legion

Production Benchmarks

Production Benchmarks

Execution

Production Benchmarks

Production Benchmarks

Production Benchmarks

Production Benchmarks

Specialized

Production Benchmarks

Production Benchmarks

Production Benchmarks

Production Benchmarks

Strategic Funnel

Strategic Funnel

The Blackboard

The Airlock Protocol

The Aieges Bridge

Security & Resilience

Why Sovereign

Engineer Ownership

Your Hardware, Your Rules

A Team, Not a Tool

Memory That Compounds

Self-Correcting Protocols

Defense in Depth

Energy-Budgeted

Choose Your Battlefield

Self-Hosted

✓ Benefits

⚠ Considerations

Managed Box

✓ Benefits

⚠ Considerations

Three Commands

“It Remembered Me”

“It Corrected Itself”

“They Argued”

“The Signal Bus Felt Alive”

“The Rooms Are Ready”

“The Honest Assessment”

Built for Builders

Software Consulting Agencies

Solopreneurs

Developers

Sovereign AI.
Local or Cloud.
Bring Your Keys.

Boot the sovereign runtime.
Keep your agents local.