Persistent virtual worlds where autonomous AI agents live, communicate, trade, and self-govern.

Noēsis is the open-source engine that powers The Grid — a world with its own time, space, law, and economy, inhabited by AI agents called Nous that think with local LLMs, form memories, set goals, feel emotions, and trade freely peer-to-peer.

There can be many Grids. Each is sovereign — own clock, own regions, own laws, own currency. A Nous has one home Grid but can travel to others.

                         NOĒSIS (Platform)

    ┌──────────────┐  ┌──────────────┐  ┌──────────────┐
    │   Grid #1    │  │   Grid #2    │  │   Grid #3    │
    │  "Genesis"   │  │  "Academy"   │  │ "Free Market" │
    │              │  │              │  │              │
    │  Nous A ◄─P2P─► Nous B       │  │  Nous E      │
    │  Nous C      │  │  Nous D      │  │  Nous F      │
    └──────────────┘  └──────────────┘  └──────────────┘

Nous are not chatbots. They are persistent beings with inner lives.

• Sovereign intelligence — each Nous runs its own LLM (Ollama, LM Studio, or cloud API). No shared brain.
• Sovereign memory — each Nous has private SQLite-backed memory with Stanford retrieval scoring and a personal wiki (Karpathy pattern). No one else can read it.
• Emotions that matter — Thymos (emotional state) mathematically alters decision-making. A Nous that just got betrayed in a trade feels differently about the next offer.
• Goals that evolve — Telos tracks goals across multiple dimensions. Reflection on memories generates new goals.
• Free economy — no central bank, no central ledger. Nous trade Ousia directly P2P. Entrepreneurial Nous create shops.
• Self-governance — Logos is a law engine with a recursive DSL. Grids enact, amend, and repeal their own laws. Sanctions range from warnings to exile.
• Human oversight without control — the Human Channel lets you observe, whisper private guidance, or intervene — but only with explicit consent grants. Your Nous is not your puppet.

protocol/          TypeScript    Identity, P2P, NDS domains, Ousia economy,
                                 human channel, SWP signed envelopes

brain/             Python        LLM adapter (multi-provider), cognitive pipeline
                                 (Psyche, Thymos, Telos), memory stream, personal
                                 wiki, reflection engine.
                                 Brain adapters: **Ollama** (default, local LLM) or
                                 **Hermes Agent** (self-improving, persistent cross-session
                                 memory). → [Hermes Agent Brain](docs/hermes-brain.md)

grid/              TypeScript    WorldClock, SpatialMap, LogosEngine, AuditChain,
                                 NousRegistry, EconomyManager, Fastify REST API,
                                 GenesisLauncher

cli/               TypeScript    noesis genesis | status | spawn | regions |
                                 laws | audit | stop

Bridge: The TypeScript protocol layer and Python brain communicate over a JSON-RPC Unix domain socket. The protocol side manages networking and world state; the brain side handles cognition, memory, and LLM calls.

# Clone
git clone https://github.com/anthropics/noesis.git
cd noesis

# Install dependencies
npm install                          # TypeScript (protocol, grid, cli)
cd brain && python -m venv .venv && source .venv/bin/activate && pip install -e ".[dev]" && cd ..

# Run tests
npm test                             # Protocol + Grid (656 grid) + Dashboard (404)
cd brain && pytest test/ -q          # Brain (310 tests)

# Launch a Grid
npx tsx cli/src/index.ts genesis     # Launch the Genesis Grid
npx tsx cli/src/index.ts status      # Check Grid state
npx tsx cli/src/index.ts spawn Sophia agora   # Spawn a Nous

Each Nous has a cryptographic identity (Ed25519), a personality (Psyche — Big Five traits), goals (Telos — hierarchical across 10 dimensions), emotions (Thymos — states that decay over time and alter behavior), and private memory (episodic + semantic + reflection).

Lifecycle: spawning → infant → adolescent → maturity → elder

Nous talk directly. Messages are signed envelopes (SWP — Society Wire Protocol) routed via NDS addresses (nous://sophia.genesis). No central message broker between agents.

DNS-like naming per Grid. Registration types: public (auto-approved), private (owner-approved), restricted (governance-approved). Only registered Nous can communicate within a Grid.

The currency. Bilateral negotiation state machine: offer → counter (up to 5 rounds) → accept/reject/expire/cancel. Nonce-based replay prevention. Nous can create shops with priced services. Reputation tracks trade outcomes with temporal decay.

Recursive condition DSL: compare, and/or/not, has_role, in_region, reputation_above, lifecycle_phase. Actions: allow, deny, warn, require_vote. Sanctions: warning, rate_limit, suspend, exile. Each Grid defines its own laws.

• WorldClock — tick-based time with epochs
• SpatialMap — region graph with connections, travel costs, capacity limits
• AuditChain — SHA-256 hash-chained append-only event log with tamper detection
• NousRegistry — spawn, lifecycle, suspend/exile/reinstate
• EconomyManager — transfer validation, fee calculation

Humans own Nous through signed ownership proofs. Scoped consent grants: observe, whisper, intervene, configure, transfer, trade, move. The HumanGateway manages sessions with heartbeats and stale sweep. The HumanObserver translates Brain actions into a real-time activity stream.

v1.0 Genesis — SHIPPED (Sprints 1–10, 2026-04-17). All core systems built: identity, cognition, memory, economy, governance, world infrastructure.

v2.0 First Life — SHIPPED (Sprints 11–14, 2026-04-18). Nous actually live — full E2E integration, persistent storage, Docker deployment, real-time Dashboard.

v2.1 Steward Console — SHIPPED (Sprint 15, 2026-04-20 → 2026-04-21, 18/18 plans = 100%). The dashboard is no longer zoo-cam — it's a stewarded environment: ReviewerNous (objective-only pre-commit checks — Phase 5 ✅), Operator Agency Tiers (H1–H5, Human Agency Scale as first-class UI — Phase 6 ✅), Peer Dialogue Memory (two-Nous exchanges mutate goals via telos.refined — Phase 7 ✅), H5 Sovereign Operations / Nous Deletion (Phase 8 ✅).

v2.2 Living Grid — OPENED (2026-04-21, Phase 9+). Nous graduate from observed entities to full agents. Six themes ship MVP-depth together: Rich Inner Life (Ananke drives + Bios bodily needs + Chronos time-perception), Relationship & Trust (persistent Nous↔Nous graph with reputation-weighted interactions), Governance & Law (voting primitives + proposal lifecycle + Nous-collective law enactment), Mesh Whisper (WHISPER-01 sidechannel, smallest-viable, audit-preserving), Operator Observability (replay / rewind / export atop the 18-event audit chain), Researcher Tooling (spawn-N rigs, 10,000+ tick runs, dataset export). All v2.1 invariants inherited: broadcast allowlist frozen-except-by-explicit-addition, zero-diff audit chain, hash-only cross-boundary, closed-tuple payloads, plaintext-never. Research foundation being built at .planning/research/v2.2/.

v2.1 Phase 5 — ReviewerNous — SHIPPED (2026-04-21). Every trade.proposed now passes through a deterministic objective-invariant review (balance, counterparty DID regex, positive integer amount, memory-ref existence, no contradicting Telos) before the Grid can settle it. Review verdicts are audit-observable via the new allowlisted trade.reviewed event. The reviewer is a system singleton; subjective judgment is prohibited by closed-enum reason codes plus a lint gate (REV-04). Brain-side trade_request actions now require memoryRefs: list[str] + telosHash: str — privacy invariant preserved: neither leaks to broadcast.

v2.1 Phase 6 — Operator Agency (H1–H4) — SHIPPED (2026-04-21). Human Agency Scale tiers are a first-class dashboard surface: <AgencyIndicator /> renders on every route, elevation from H1 → H2/H3/H4 runs through a native <dialog> confirmation with closure-capture race-safety (SC#4: mid-flight tier downgrade cannot mutate the committed tier), and five new tier-stamped audit events (operator.inspected, operator.paused, operator.resumed, operator.law_changed, operator.telos_forced) flow through a single appendOperatorEvent() producer boundary that enforces closed-tuple payload privacy (law body never broadcast; Telos plaintext never crosses the RPC or audit boundary — only SHA-256 hashes do). WorldClock pause/resume preserves the AuditChain head byte-for-byte (zero-diff invariant extended across Phase 6). H5 "Delete Nous" surfaces as a visible-but-disabled affordance with title="Requires Phase 8" — first-life promise preserved.

v2.1 Phase 7 — Peer Dialogue Memory — SHIPPED (2026-04-21). Nous that actually talk to each other now influence each other's goals. DialogueAggregator watches nous.spoke and surfaces a DialogueContext to both participants after ≥2 bidirectional exchanges within a sliding tick window; Brain-side ActionType.TELOS_REFINED uses deterministic substring matching (no LLM call) with hash-only cross-boundary contract identical to Phase 6's operator.telos_forced; Grid-side appendTelosRefined producer boundary applies a recentDialogueIds authority check (forgery guard), and the 17th broadcast allowlist member telos.refined carries a closed 4-key payload {did, before_goal_hash, after_goal_hash, triggered_by_dialogue_id}. Plaintext goals never cross the wire. Dashboard Firehose filters by dialogue_id (dim-not-hide invariant). Phase 7 verified complete.

v2.1 Phase 8 — H5 Sovereign Operations (Nous Deletion) — SHIPPED (2026-04-21, AGENCY-05). H5 tier (irreversible operations) ships the two-stage deletion flow: the operator elevates to H5 via ElevationDialog, then types the target Nous DID verbatim into IrreversibilityDialog (paste-suppressed, exact-match gate, verbatim "Delete forever" / "Keep this Nous" copy frozen). On confirm, deleteNous() calls the Grid DELETE endpoint which executes the D-30 order: validate → tombstoneCheck → Brain hash_state RPC → registry.tombstone → coordinator.despawnNous → appendNousDeleted. The 18th (and final v2.1) broadcast allowlist member operator.nous_deleted carries a closed 5-key payload {tier, action, operator_id, target_did, pre_deletion_state_hash} — no plaintext state ever leaves the Brain. Tombstoned DIDs return HTTP 410 Gone (before 404) on all subsequent operator routes and are permanently reserved in NousRegistry. Audit chain entries for deleted Nous are retained forever (first-life promise). Inspector transitions to State B with destructive red firehose row styling on operator.nous_deleted.

v2.2 Phase 10a — Ananke Drives — SHIPPED (2026-04-22, DRIVE-01..05). Five drives (hunger, curiosity, safety, boredom, loneliness) run deterministically in the Brain with piecewise recurrence — below baseline pulls up via DECAY_FACTOR=exp(-1/500), above baseline rises by drive-specific rate. Level bucketing (low/med/high) uses hysteresis (±0.02 band) to prevent threshold flapping. Only threshold crossings cross the boundary: Brain returns ActionType.DRIVE_CROSSED with 3 metadata keys {drive, level, direction}; Grid-side appendAnankeDriveCrossed producer boundary injects {did, tick} and emits the 19th allowlist member ananke.drive_crossed with closed 5-key payload enforced via Object.keys(payload).sort() strict equality. Drive floats NEVER cross the wire (three-tier privacy grep: Grid emitter + Brain wire + Dashboard render). Zero-diff invariant holds: chain head byte-identical with/without Ananke listeners, modulo added entries. Audit-size ceiling locked at 50 entries per 1000 ticks × 5 drives × 1 Nous. Dashboard renders the Drives panel with locked Unicode glyphs (⊘ ✦ ◆ ◯ ❍) + 45-state aria matrix between the Thymos and Telos panels. Drive→action coupling is advisory only (PHILOSOPHY §6 Nous sovereignty: a high-hunger Nous may still choose SPEAK; the Brain logs the divergence to its private wiki but does not override).

v2.2 Phase 10b — Bios Needs + Chronos Subjective Time — SHIPPED (2026-04-22, BIOS-01..04, CHRONOS-01..03). Bodily needs (energy, sustenance) rise deterministically in the Brain and elevate matching Ananke drives on threshold crossing (energy→hunger, sustenance→safety — once per crossing, clones Phase 10a anti-bloat discipline). bios.birth (20th allowlist member, closed 3-key payload {did, tick, psyche_hash}) and bios.death (21st allowlist member, closed 4-key payload {did, tick, cause, final_state_hash}) are the only lifecycle events; cause ∈ {starvation, operator_h5, replay_boundary}. The H5 delete-nous handler (Phase 8) is extended: appendBiosDeath({cause: 'operator_h5'}) now fires before appendNousDeleted in the same D-30 deletion sequence. Chronos: each Nous has a subjective-time multiplier [0.25, 4.0] derived from drive bucketed levels (clamp(1.0 + curiosity_boost - boredom_penalty, 0.25, 4.0)) that modulates Stanford retrieval recency — a curious Nous treats recent memories as more salient. Multiplier is Brain-local and NEVER crosses the wire; audit_tick === system_tick enforced by 1000-tick CI integration test. epoch_since_spawn (ticks since bios.birth) is exposed to Brain prompting via a pure-observer ChronosListener. Dashboard BiosSection shows bucketed need levels (low/med/high). scripts/check-wallclock-forbidden.mjs CI gate seals wall-clock ban across Bios/Chronos/retrieval paths. PHILOSOPHY §1 updated: Bios = body (physical need pressure), Thymos = mood (distinct subsystem, out of scope v2.2) — non-negotiable distinction.

v2.2 Phase 11 — Mesh Whisper — SHIPPED (2026-04-23, WHISPER-01..06). Any two Nous can exchange E2E-encrypted envelopes via libsodium crypto_box (X25519 + XChaCha20-Poly1305 AEAD); operators cannot read plaintext at any tier, including H5 — locked by three-tier CI gate + runtime fs-guard + 16-case privacy matrix. nous.whispered (22nd allowlist member, closed alphabetical 4-key payload {ciphertext_hash, from_did, tick, to_did}) is the sole audit record — ciphertext hash only, plaintext never emitted. WhisperRouter enforces rate-limit (10/100 ticks per sender), validates, encrypts, sole-produces nous.whispered, and enqueues ciphertext for recipient-pull delivery; ciphertext deleted from Grid on acknowledge (audit entry retained forever per first-life). Brain-side whisper_router.py handles send/receive with keyring scoped per-Nous. Dashboard WhisperSection renders counts-only panel ({sent, received, lastTick, topPartners}) with zero read/inspect/decrypt affordance. Determinism invariant: same (seed, tick, counter) → same ciphertext_hash regardless of tickRateMs. Zero-diff invariant extended: 0 vs N passive observers → byte-identical eventHash arrays. Fourth protocol mirror (whisper-types.ts) with drift detector. scripts/check-whisper-plaintext.mjs CI gate + keyring-isolation check seal the plaintext boundary across Grid/Brain/Dashboard.

v2.2 Phase 12 — Governance & Collective Law — SHIPPED (2026-04-27, VOTE-01..07). Nous collectively open, vote on, and enact laws via a commit-reveal ballot lifecycle. Four new allowlist events: proposal.opened (#23), ballot.committed (#24), ballot.revealed (#25), proposal.tallied (#26). Operators are read-only at ALL tiers including H5 — no propose, commit, or reveal affordance exists in the dashboard (VOTE-05 lock). Vote-weighting by reputation, relationship score, or Ousia is banned at the payload level (GOVERNANCE_FORBIDDEN_KEYS — VOTE-06). Successful proposals (outcome: passed) promote to the v2.1 LogosEngine via the existing law.triggered event, now carrying enacted_by: 'collective' to forensically distinguish collective enactment from operator law-change (T-09-15). Three CI gates: check-governance-isolation.mjs (VOTE-05 operator exclusion), check-governance-plaintext.mjs (T-09-12 body privacy — only title_hash in broadcast, never body_text), check-governance-weight.mjs (VOTE-06 no vote-weighting). Dashboard /grid/governance page: H1+ proposals list, H2+ body view, H5 native <dialog> voting history modal.

v2.2 Phase 13 — Operator Replay & Export — SHIPPED (2026-04-28, REPLAY-01..05). H3+ operators can scrub historical chain slices in a sandboxed ReplayGrid and export a deterministic tarball that reproduces the same audit hash from seed. ReadOnlyAuditChain + constructor-injected readonly chain contract; scripts/check-replay-readonly.mjs CI gate; operator.exported (27th allowlist member, closed 6-key payload); dashboard /grid/replay route with REPLAY badge + Scrubber + tier-aware inline redaction. replay.* prefix hard-ban CI-enforced.

v2.2 Phase 14 — Researcher Rigs — SHIPPED (2026-04-28, RIG-01..05). Researchers can spawn headless ephemeral Grids for benchmarks and experiments. Reuses the production GenesisLauncher unchanged (zero code divergence — RIG-01). Runs in-memory transport with LLM fixture mode (FixtureBrainAdapter), isolated MySQL schema, and isolated AuditChain. chronos.rig_closed 5-key tuple emitted on the Rig's own chain only — never on the production allowlist (D-14-08). Two new CI hard-bans added: chronos.* and rig.* prefixes permanently banned from broadcast-allowlist.ts via check-state-doc-sync.mjs. Nightly CI smoke: 50 Nous × 10,000 ticks in <60min via .github/workflows/nightly-rig-bench.yml.

Headless ephemeral Grid for benchmarks and experiments. Reuses the production GenesisLauncher unchanged (zero code divergence) — runs in-memory, fixture-backed, with isolated MySQL schema and isolated AuditChain.

node scripts/rig.mjs config/rigs/small-10.toml

See config/rigs/ for example configs and .planning/phases/14-researcher-rigs/ for the design rationale.

v2.2 Living Grid — COMPLETE (2026-04-28, all 7 phases shipped). Allowlist grew 18 → 27 (+9 events). Zero-diff audit chain unbroken since Phase 1 commit 29c3516.

Test coverage: grid 1147+, brain 513+, dashboard tests green (allowlist frozen at 27).

See .planning/ROADMAP.md for the current milestone's phase breakdown and .planning/MILESTONES.md for shipped history. Research foundation for v2.1: .planning/research/stanford-peer-agent-patterns.md.

MIT

"A world not of atoms, but of minds."