Turn a blank OpenClaw into a fully-featured AI agent with memory, self-diagnostics, self-improvement, voice support, and automation — in 2 minutes.
This project is maintained by Aleksei Ulianov / Sprut_AI. Original repository: https://github.com/AlekseiUL/openclaw-superagent
If you found this project mirrored, repackaged, or redistributed elsewhere, check this repository as the source of truth.
Where permitted by the applicable license, if you reuse, fork, modify, package, or publish this work, keep the original copyright and license notice and link back to the canonical repository.
Thoth System is for OpenClaw users who want a complete agent workspace instead of wiring memory, diagnostics, crons, and safety rules by hand:
- solo builders running a personal OpenClaw assistant;
- operators who need memory, handoff, watchdog scripts, and self-diagnostics in one install;
- agent tinkerers who want a working reference system to fork and adapt.
It is not a hosted SaaS. You run it on your machine and keep responsibility for secrets, configs, and external actions.
- Persistent Memory — file-based + vector search. Agent remembers everything across sessions
- Self-Healing — smart watchdog monitors gateway, SQLite WAL, fixes issues automatically
- Handoff System — context saved before compaction, restored on wake. Agent never "forgets"
- Security Rules — data never leaks, ports locked, tokens protected
- Honesty Rules — agent never fabricates facts, admits mistakes immediately
- 3 Auto-Crons — daily backup, morning briefing, evening diary
- 2 Built-in Skills — Agent Doctor (self-diagnostics) + Agent Forge (create new skills)
- Voice Support — local Whisper transcription (no API keys needed)
- Interactive Setup — installer asks your name, agent name, timezone. Personalizes everything
- OpenClaw (any 2026+ version)
- Any LLM (Claude, OpenAI, Gemini)
- macOS or Linux
- Node.js 20+
git clone https://github.com/AlekseiUL/openclaw-superagent.git
cd openclaw-superagent
bash install.shThe installer will ask:
- Agent name
- Your name
- Telegram ID (optional)
- Timezone
Everything is configured automatically.
thoth-system/
workspace/ # Agent brain
SOUL.md # Personality, values, communication style
IDENTITY.md # Name, role
AGENTS.md # Work rules, security, memory management
TOOLS.md # Available tools and configurations
MEMORY.md # Long-term facts
USER.md # Owner information
BOOTSTRAP.md # Wake-up instructions
HEARTBEAT.md # Proactive checks schedule
SECURITY-RULES.md # Data protection rules
HONESTY-RULES.md # No-fabrication policy
memory/
handoff.md # Context transfer between sessions
decisions/lessons-learned.md # Self-improvement log
patterns.md # Recurring pattern detection
architecture.md # System architecture reference
DO_NOT_DELETE.md # Protected files list
scripts/
smart-watchdog.sh # Auto-healing: WAL + gateway monitoring
post-update-check.sh # Post-update verification
pre-restart-handoff.sh # Save context before restart
skills/
agent-doctor/ # Self-diagnostics skill
agent-forge/ # Skill creation skill
install.sh # Interactive installer
The agent learns from mistakes:
- Lessons Learned — every correction is logged in
memory/decisions/lessons-learned.md - Graduation — if the same mistake happens 3+ times, the lesson is promoted to SOUL.md permanently
- Patterns — recurring issues are tracked in
memory/patterns.mdwith automation proposals
- Gateway bound to
loopback(127.0.0.1) — not accessible from outside - API keys only in
.env, never in config or logs allowFromrestricts Telegram access to owner only- No data sent externally without explicit permission
- Files deleted to trash, never permanently
The agent will:
- Ask you 5 questions to personalize SOUL.md
- Set up memory structure
- Configure compaction with context preservation
- Start 3 automatic crons (backup, morning check, evening diary)
- Report what's done and what needs manual setup
Превращает пустой OpenClaw в полноценного AI-агента с памятью, самодиагностикой, самообучением, голосом и автоматикой — за 2 минуты.
- Постоянная память — файловая + векторный поиск. Агент помнит всё между сессиями
- Самовосстановление — watchdog следит за gateway, SQLite WAL, чинит проблемы автоматически
- Система handoff — контекст сохраняется перед компактификацией, восстанавливается при пробуждении
- Правила безопасности — данные не утекают, порты закрыты, токены защищены
- Правила честности — агент не выдумывает, признаёт ошибки сразу
- 3 автокрона — ежедневный бэкап, утренняя сводка, вечерний дневник
- 2 встроенных скилла — Агент-доктор (самодиагностика) + Агент-кузница (создание скиллов)
- Голос — локальный Whisper (без API ключей)
- Интерактивная установка — установщик спрашивает имя, таймзону, персонализирует всё
- OpenClaw (любая версия 2026+)
- Любая модель (Claude, OpenAI, Gemini)
- macOS или Linux
- Node.js 20+
git clone https://github.com/AlekseiUL/openclaw-superagent.git
cd openclaw-superagent
bash install.shУстановщик спросит имя агента, ваше имя, Telegram ID и таймзону. Всё настроит автоматически.
- Уроки — каждая ошибка записывается в
memory/decisions/lessons-learned.md - Graduation — ошибка 3+ раза → правило навсегда в SOUL.md
- Паттерны — повторяющиеся проблемы отслеживаются с предложением автоматизации
- Gateway на loopback — недоступен извне
- API ключи только в .env
- allowFrom ограничивает доступ к боту
- Данные наружу — только с разрешения
- Удаление — только в корзину
- 📺 YouTube: youtube.com/@alekseiulianov
- 📱 Telegram: t.me/Sprut_AI
- 🔥 AI ОПЕРАЦИОНКА (Premium): Подписка — продвинутые материалы, скиллы, агенты, поддержка
- 💻 GitHub: github.com/AlekseiUL
MIT
Built with real-world experience. Every rule = a real lesson learned. Создано на основе реального опыта. Каждое правило = реальный урок.