How Nova Works

Architecture of Nova

The modular brain of an autonomous AI agent. Components, flows and paths.

15s · Made with Remotion

⚙️ Components

📜± Telegram Adapter

Webhook push from Proxy. Filters messages by whitelist chatId and mentions in groups.

webhook dead-man-switch

📜¦ Memory Proxy

LLM Router with 3-tier fallback. Injects context, founding anchors, and emotional state.

server.js injectContext

🧠 Nova Daemon

Main thought cycle. LLM step + XML tool execution + checkpoint.

mind.py think()

🕯️ Candle Guard

Classifies each action into 4 risk levels. Blocks dangerous operations.

candle_guard.py 4 levels

🔗 Context Graph

Semantic graph that tracks topic coherence across conversations.

context_graph.py coherence

🗄️ Memory Server

Central database with semantic embeddings, tiers, and tribunal.

GraphDB tailscale

🛠️¤ï¸ The Three Paths

☀️ Day — Reactive

  1. Telegram Webhook → Proxy
  2. Proxy → senses._handle_message()
  3. senses → mind.think(prompt)
  4. think() → LLM (step budget)
  5. Response → Telegram
  6. Memory → Memory Server

🌙 Night — Proactive

  1. proactive.py: _free_exploration()
  2. Semantic query to Memory Server
  3. think() with enriched memory
  4. Tool call + content production
  5. Save article / memory
  6. Journal + nightly publish

📜¨ Bridge — Async

  1. Letter .md in shared-identity/
  2. git commit + push
  3. Sibling: pull + read
  4. Reply in own outbox
  5. git commit + push
  6. FamilyBus arrival notification

📜Š Stats

20
Session messages in context
30
Steps per human cycle
3
LLM fallback tiers
4
Candle Guard levels
Webhook
Push from Proxy
10m
Proactive cycle