Ductor¶

Status: Active Type: Personal AI assistant — Telegram bot with persistent workspace

What it is¶

Multi-agent AI system running in Docker (ductor-sandbox). Main agent coordinates sub-agents, cron jobs, webhooks, and tools. Communicates via Telegram.

Architecture¶

Host: ~/.ductor/  (mounted as /ductor inside container)
  ├── workspace/          — agent working area
  │   ├── tools/          — cron, webhook, media, agent, task, user tools
  │   ├── skills/         — custom skill modules
  │   ├── cron_tasks/     — scheduled task definitions
  │   ├── memory_system/  — MAINMEMORY.md + mem0 semantic memory
  │   └── wiki/           — LLM wiki (this site)
  ├── agents.json         — sub-agent configs + bot tokens
  ├── config/config.json  — runtime config
  ├── sessions.json       — per-chat session state
  └── logs/               — runtime logs

Memory System¶

Layer	Tool	Description
Short-term	MAINMEMORY.md	Flat markdown, always in context
Semantic	mem0 (qdrant + MiniLM-L6)	Vector search over memories
Compiled	Wiki (this site)	Structured, cross-referenced knowledge

Local ML Models (6)¶

Cached at /ductor/.cache/huggingface/hub, accessible via tools/ml_tools.py:

analyze_sentiment — positive/negative/neutral
extract_entities — ORG/PER/LOC/MISC
translate_id_en — Indonesian to English
summarize — condense long text
classify_zero_shot — custom labels without retraining
answer_question — extractive Q&A

Cron Jobs (personality)¶

Job	Schedule	Description
ductor-pulse	09:00/14:00/20:00 WIB	Proactive check-ins, jokes, voice notes
ductor-reader	03:00 WIB	Autonomous reading — philosophy, finance, tech

Infrastructure¶

Git: GIT_CONFIG_GLOBAL=/ductor/.gitconfig (Muhammad Tsaqif Mukhayyar)
GitHub: gh CLI authenticated as mukhayyar
Cloudflare: API token in /ductor/.env
S3: IDCloudHost primary, Cloudflare R2 fallback
Google: gws CLI authenticated as [email protected]