Orreth

Flagship

A Memory Substrate for Governed AI Universes — every actor an identity, every memory signed, nothing unattributable

Overview

An agent today is only as good as the data in its context window, and its data is erased on reboot. Orreth inverts the model: the running system is the memory. Universes of Ecosystems of Fields of Agents — one recursive primitive where the tier is just a profile — with memory keyed to permanent identities, signed at creation, distilled as it rises, and retrievable across the whole estate under strict authorization. Policy cascades down and can only tighten; memory rises up and is pruned; retrieval escalates by time-horizon, sourced and verified or not at all.

On top of that substrate, the organization runs itself: resident agents staff every layer (an identity agent minting every credential, a librarian governing knowledge, a farm keeper attesting every tool, a wrangler pinning every model deal, a smith proposing every improvement), a workforce carries the objectives humans set, and every consequential change waits at a human gate. It is agentic end to end — and auditable at every hop.

33
Design Dives (0000–0032)
132
Conformance Tests
1
Binary — Tier as a Profile
3
Tiers on One Laptop
The End of the Context Window — an enterprise's memory across spacetime, with a context window drawn to scale

An enterprise's memory across spacetime — tier horizons, distillation arrows, one question riding from a working seat up through all of time. In the corner: a context window, drawn to scale. The context window is a porthole; this is the ocean.

Core Concepts

Identity Is the Thread

Every agent, tool, and mind is a permanent DID-anchored identity. The process is disposable; the identity and its memory are not. Reboot is not death — the same self re-joins through a governed gate.

Memory Rises & Distills

Fields keep raw detail ~90 days, ecosystems a year, the apex forever — distilled, with signed provenance chains back to source. Memory changes resolution, never existence. Years stay affordable.

The Human's Ladder

Objective · Intention · Observation · Thought. Humans alone originate Objectives; plans wait for human approval before anything fans; each seat receives one intent, a budget, and citation refs — never the plan, never its siblings.

Governed Self-Improvement

Prompts, skills, and workflows are versioned assets. A standing improver proposes new versions from evidence; governance grades the change by diff; the human adopts or declines at a gate. Siblings, never silent successors.

The Spacetime Window

Stand at any moment of the world's history and see it whole — every record signed, every render a governed, tokened query. Open any finished objective and walk to any seat: what it knew, what it cost, who graded it.

Supply-Chain Doors

Tools and models are identities with hash-pinned manifests. A changed manifest quarantines instantly (the rug-pull door — CVE-2025-54136's move, refused structurally) and drops dependent knowledge to 'investigating'. Discredit a source and a recall walk marks its whole lineage.

Architecture — Three Flows


  ┌─────────────────────────────────────────────────────────────┐
  │  UNIVERSE — foundation & apex · all-time memory, distilled  │
  │  residents: identity root · librarian · smith · governance  │
  └───────────────┬─────────────────────────────▲───────────────┘
                  │ policy cascades DOWN        │ memory rises UP
                  │ floors never loosen         │ pruned & distilled
  ┌───────────────▼─────────────────────────────┴───────────────┐
  │  ECOSYSTEMS — lines of business · tenant-isolated worlds    │
  └───────────────┬─────────────────────────────▲───────────────┘
                  │                             │
  ┌───────────────▼─────────────────────────────┴───────────────┐
  │  FIELDS — where the work happens · ~90-day raw horizon      │
  │  workforce agents: living identities, joined at a gate      │
  └─────────────────────────────────────────────────────────────┘

  retrieval escalates UP by time-horizon · Sourced + Verified
  every record Ed25519-signed · every render a governed query

Technology Stack

Plane

Rust
Axum
PostgreSQL 16
Object Store
Docker Compose

Reference

Python 3.12
132-test conformance suite
Byte-identical canonicalization

Identity & Trust

Ed25519
DID (did:web / did:key)
Capability tokens
Content addressing (SHA-256)

Cognition

LiteLLM
Governed model plane
Per-agent metering
Budget floors

Governance

HITL quorums
Severity lanes
Tighten-only policy cascade
Signed markers

Glass

The Console (served by the daemon)
Live orrery
Spacetime window
Objective ledger

The Lineage

Everything else on this site was practice for Orreth

CortexObserver proved a governed AI workforce commanded in plain English. ProblemSolver's Thinker–Planner–Critic–Judge pipeline became the ancestor of Orreth's agent chassis. MCPFarm's centralized tool governance became the Tool Farm — where every tool is an identity with a worldline and a hash-pinned manifest. Each platform answered a question; together they raised the one Orreth answers: what if the workforce, its tools, its knowledge, and its history all lived in one governed, signed, permanent world?

Orreth is that world — designed in 33 public-methodology dives with every decision locked on a ledger, built twice (a Python reference and a Rust plane, conformance-tested against each other byte for byte), and run live at demo.orreth.ai.

Implementation Highlights

One Binary, Any Tier

orrethd runs a Universe, an Ecosystem, or a Field — the tier is just a profile. A full three-tier universe stands up on a laptop with one command, each layer pulling its policy floors from the one above at boot. Multi-tenant isolation, trust-root pinning, and retrieval that escalates deep-time questions up the tree.

Nothing Grades Its Own Yardstick

Every cycle of agent thought lands as a scribe-signed run record — author never equals agent — pinned to the exact policy that governed it. "The agents are improving" is a measurable claim about signed history, not a vibe.

A Recall — for Knowledge

External knowledge enters quarantined at 0.0000 confidence through identified sources and earns promotion with receipts. Discredit a source and every entry it fed — plus everything derived from those, however deep — is enumerated and marked recalled. Annotated, never rewritten. Chain-of-custody discipline for a training corpus.

The Human Holds the High Lane

100% of change is graded into severity lanes: low auto-approves on the record, medium notifies, high waits for a human. Approvals are quorum-based where it matters most — below the quorum, the gravest actions aren't forbidden, they're structurally unavailable.