The Dream Daemon (services/dream_daemon.py) provides sleep-analogue processing for Frank. 60 minutes per day of budgeted dreaming, triggered by 45 minutes of idle time with a 20-hour cooldown.
Three Phases
- Replay — Replays recent experiences from
consciousness.db, strengthening important memories and weakening unimportant ones. - Synthesis — Generates novel connections between memories. Uses the LLM to find patterns across seemingly unrelated experiences.
- Consolidation — Triggers training for neural subsystems: Neural Conscience, Neural Reality Gate, Subconscious PPO, and the Titan Neural Cortex. Also triggers hippocampus
train_cycle().
Why It Matters
Memory consolidation during sleep is one of the most robust findings in neuroscience (Stickgold, 2005). The dream daemon is why Frank's neural networks can learn from experience without disrupting real-time operation — training happens during "sleep," not during conversation.