🫀 Zoetron

🇩🇪 Deutsche Version

An autonomous agent organism that gets smarter from every experience — running on free OpenRouter models, with an AutoRouter immune to model mortality.

Zoetron is not an agent framework. It is an organism: heartbeat, curiosity, sleep, a world model, a self-model, and an immune system against dead models.

🧠 Watch its memory live: zoetron landing page — its brain, ideas board and memory stream on one page, redrawn on every heartbeat · 2D knowledge graph · 3D brain — every fact it learns becomes a neuron; age fades the glow.

📈 Watch it get better — or not: football learning curve. The organism predicts match outcomes from 153,000 games (2000–2024) and is scored by the actual results, not by a language model. Baseline to beat: 50.2 % on unseen matches, without bookmaker odds; the bookmaker gets 51.5 %. A third set of 13,143 games is sealed and never touches the organism's machine — because a test set answered a hundred times stops being a test. The page itself is a React/Vite/Tailwind/GSAP app in site/ that builds into docs/.

Goal ──► PLAN ──► SIMULATE ──► SELF-ASSESS ──► BUILD ──► CRITIQUE
          ▲                                              │
          └────────── score < 8: next iteration ◄────────┘
                     ▼
     JSONL memory (facts + events) → DREAM distills insights
                     ▼
     DRIVE generates the next self-chosen goals from knowledge gaps

The Organs

Organ	Nature analog	File	What it does
💓 HEART	heartbeat	heartbeat_daemon.sh	every 30 min (48/day): REFLEX → PRUNE → DREAM → DRIVE → ACT (+HANDS) → FOOTBALL SCORING → WILL → GENOME → TEACH → SELF-LEARN → TELEGRAM → IDEAS → LEADERBOARD → GRAPH+3D-BRAIN → SELF-DIAGNOSIS → RETRO → git/wiki sync
🧬 GENOME	innate DNA	genome.py	5 instinct rules (data/genome.json) injected into every LLM prompt — protect memory, creator first, self-preservation watchdog, honesty, learn from every failure
🔥 DRIVE	curiosity/drive	drive.py	generates its own goals from memory gaps — in plain German, with a human-readable "why"
💭 DREAM	sleep/replay	dream.py	distills experiences into lasting insights + skill proposals
🪞 METACOG	self-awareness	metacog.py	predicts its own scores, measures calibration
🌍 SIMULATE	world model	simulate.py	dry-runs plans before acting on them
👁 SENSES	eyes/ears	senses.py	watches HN frontier + web search; fresh signals feed DRIVE
🔬 RESEARCHER	scout	swarm role 4	fresh web facts before planning (Phase 2 of ROLES)
🕸 SEMANTIC MEMORY	meaning cortex	semantic.py	TF-IDF + synonym fields + cross-domain bridging: finds related knowledge by meaning, not wording — injected into every plan
✋ HANDS	motor system	hands.py	executes artifact code in a sandbox; successes become reusable tools (data/tools/)
🧬 EVOLUTION	reproduction	evolution.py	N solution variants → critic picks the fittest; winners are inherited, losers become forbidden anti-patterns (never repeated)
🩸 METABOLISM	stress axis	metabolism.py	rate limits & timeouts raise stress → smaller plans; cortisol decay
✂️ PRUNE	apoptosis	prune.py	archives stale memories, keeps recalled facts, insights are immortal
🎯 WILL	prefrontal cortex	will.py	long-term missions with measurable targets survive across cycles; they outrank short-term curiosity (whisper still wins)
🪞 RETRO	conscience	retro.py	reviews its own cycles after every heartbeat: honest findings become retro_lesson: facts that feed the next planner prompt
🕸 GRAPH+BRAIN	associative cortex	graph.py, graph_view.py, brain_view.py	live knowledge graph (2D physics view + 3D brain with hemispheres & synapse pulses), semantic linking: synapses form by TF-IDF meaning (cosine ≥ 0.18), not word overlap — "timeout during build" finds "aborted due to time exceeded"; duplicates consolidated automatically
💡 IDEAS BOARD	daydream notebook	ideen.py	collects its own autonomous ideas (skill wishes, self-chosen goals, dream insights) into a public board — refreshed every heartbeat
⚡ REFLEX PATH	spinal cord	ACT-first logic	checks the toolbelt before the LLM thinks — token-free runs for known work; 6 h throttle per goal+tool combo (no spam)
🛡 CRITIC	adaptive immunity	swarm.py	adversarial quality gate (score 0–10); demands runnable code, not prose
💬 WHISPER	human voice	data/fluester_goals.json	human priority goals jump the queue (logged as drive_whisper)
🧬 INHERITANCE	heredity	vererbung.py	hands the next attempt its own best measured submission plus the instruction change exactly one thing — before this, every run started from the same toy scaffold and nothing was ever inherited
🩺 SELF-DIAGNOSIS	pain receptors	selbstdiagnose.py	parses the heartbeat log for crashes, timeouts and tracebacks (organ, error class, file:line) and files them as memory events — a crashed phase used to be absent to RETRO, not failed
🛤 PATHWAYS	Hebbian synapses	bahnen.py	persistent weighted edge table with a dopamine gate: only runs that produced runnable code strengthen a connection; unconfirmed links decay to nothing in 4 days
😴 REPLAY	sleep consolidation	nachspiel.py	in DREAM, re-scores earlier traces once the external measurement has moved — the critic alone is never allowed to reinforce a pathway
⚖️ EXTERNAL PROOF	incorruptible judge	externer_beweis.py	a goal can be closed only by a measurement outside the system; if the critic says converged and the measurement disagrees, the measurement wins
🧯 CONTRADICTION GUARD	immune tolerance	widerspruch.py	archives tool-call noise and key collisions in PRUNE — archived, never deleted
🔌 ACCOUNT LOCK	hibernation	kontosperre.py	detects a rate-limited account (needs all candidates failing on account errors, 3+ pieces of evidence), sleeps until the reset time, wakes itself
🧱 SHIELD	blood-brain barrier	schutzwall.py	guards os.remove/rmtree/rename outside the sandbox for executed snippets — protection against accident, explicitly not a security boundary
🚦 TOOL ADMISSION	quality control	werkzeugpruefung.py	five rules decide what may enter the toolbox; rejects go to data/tools_abgelehnt/, never to the bin
🩻 AUTOROUTER	immune system	router.py, model_health.py	ranks free models by measured failure rate and latency, not by metadata; a model that never answers no longer holds a slot
🧠 Memory	hippocampus	memory.py	JSONL facts + event log + keyword recall; auto-deduplicated

The Autonomous Loop

Every 30 minutes (48 cycles/day), Zoetron runs a full developer cycle without human input — and thinks about itself while doing it. Telegram reports arrive once per beat so the thinking never spams the human:

HERZSCHLAG (every 30 min)
 ├─ ⚡ REFLEX   toolbelt first — known work runs token-free (6 h throttle)
 ├─ ✂️ PRUNE    archive stale memories
 ├─ 💭 DREAM    distill new insights + skill ideas from recent experience
 ├─ 🔥 DRIVE    pick its own next goal (curiosity) or a whispered one (you)
 ├─ 🐜 ACT      swarm plans → builds → critic scores 0–10
 │     score ≥ 8: issue closes itself (status:erledigt) ✅
 │     score < 8: evolution breeds better strategies, losers become
 │                forbidden anti-patterns, winners are inherited
 ├─ ✋ HANDS    artifact code executed in a sandbox; successes become tools
 ├─ 📱 TELEGRAM posts the proof of execution (every beat)
 ├─ ⚽ SCORING  an external judge grades football submissions — no LLM involved
 ├─ 🕸 GRAPH    rebuilds the live knowledge graph — semantic synapses
 ├─ 💡 IDEAS    refreshes the public ideas board (its own inventions!)
 ├─ 🩺 DIAGNOSE reads its own log for crashes and files them as events
 └─ 🪞 RETRO    reviews its own cycles: "what went well / badly?"
               findings are stored as lessons for future planners

• Scoring: 0–10 from an adversarial critic LLM; 8+ = converged. Details: Wiki → Scoring
• Whisper channel: humans can inject priority goals via data/fluester_goals.json — Zoetron executes them before its own curiosity and logs the intervention as drive_whisper.
• Public self-management: issues, label kanban, releases, wiki and the landing page are maintained by the organism itself.

💡 What it thinks about every 30 minutes

Every heartbeat generates fresh ideas you can watch grow in the open:

Phase	Typical ideas coming out of it
💭 DREAM	distilled insights ("what did I learn since last beat?") + concrete skill proposals for new abilities
🔥 DRIVE	a brand-new goal with a human-readable German why — chosen because memory has a gap there
🐜 ACT	runnable Python artifacts per task (code is mandatory, prose alone scores low)
🧬 EVOLUTION	up to 3 fundamentally different attack angles per failure; the fittest becomes inherited strategy
🪞 RETRO	honest self-findings like "average score only 1.5/10 – code mandate still too weak" or "N hands failed – build smaller standalone scripts"

Recent real examples from the memory: "Redesign MemoryStore as self-pruning ring buffer", "Embedding-driven organ integration: replace BFS subgraph extraction with learned similarity", "Enforce MemoryStore contract validator before each hand-action batch; fail fast."

Source Access

The source code is not public (locked area, all rights reserved). This repository is Zoetron's public logbook: architecture, wiki, live memory, knowledge graph, 3D brain and weekly reports — everything the organism writes about itself, refreshed every heartbeat. There is nothing to install here.

License & collaboration only on a serious, written request via GitHub @djlex83 — with real name, background and a concrete purpose. Licensing requires a contract + NDA.

CLI reference (for license holders)

python -m zoetron.cli run "Your goal" --swarm --save-artifacts out/
python -m zoetron.cli models   # inspect the live routing queue
python -m zoetron.cli drive    # let it generate its own goals
python -m zoetron.cli dream    # sleep phase: consolidate experience
python -m zoetron.cli status   # self-model & calibration

Configured via .env: OPENROUTER_API_KEY and ZOETRON_MODEL=auto (AutoRouter = model immortality). Without an API key: deterministic mock brain, all 168 tests pass fully offline. ZOETRON_MODEL=auto automatically discovers every free tool-capable OpenRouter model and learns from its own runs which ones perform well.

Why This Exists

Today's LLM agents are oracles on demand: frozen weights, no goals of their own, no sleep, no self-knowledge. Zoetron builds the missing organs in software — a self-hosted research organism for continuously learning systems.

Status

Early prototype (v0.x), active development, design changes fast. First live runs showed: the critic catches real product inconsistencies, DREAM diagnoses pipeline bugs from its own experience, and the AutoRouter skips dead models automatically. Milestone: the first fully autonomous goal converged (score 8/10) and closed its own GitHub issue — the complete think → act → assess → close loop now runs unattended on every heartbeat.

Reality check, 23 Aug 2026. A thorough audit found the organism had been recording attempts as successes: of 22 goals reported done, one had actually been achieved, and of 31 registered tools, three did anything. Measurement, tool admission and logging are repaired — convergence must now be scored, tools must pass a four-rule admission check, and aborts are no longer swallowed. The full account, including the numbers we would rather not publish: Werkstattbericht 23.08.2026.

🔒 Source Code Notice / Quellcode-Hinweis

🇩🇪 Zugang zum Quellcode: Der Code ist nicht öffentlich (gesperrter Bereich). Lizenz & Zusammenarbeit nur auf seriöse, schriftliche Anfrage über GitHub @djlex83 – mit echtem Namen, Hintergrund und konkretem Anliegen. Formlose Massen-Nachrichten, Werbeangebote und „Preis?“-One-Liner werden ignoriert. Lizenzvergabe erfordert Vertrag + NDA. Alle Rechte vorbehalten.

🇬🇧 Source access: The code is not public (locked area). License & collaboration only via serious written request at GitHub @djlex83 — real name, background and a concrete matter required. Generic mass messages, sales offers and one-line “price?” inquiries are ignored. Licensing requires contract + NDA. All rights reserved.

License

Alle Rechte vorbehalten – Nutzung/Kopieren/Weitergabe nur mit Genehmigung des Autors. Siehe LICENSE. Anfragen: Issue oder @djlex83.