Brain Persistence — Save & Load Complete Brain State
The brain persistence system serializes the ENTIRE cognitive state of a creature — SNN weights, world model, emotions, memories, concept graph, and all module states — into a single .pt file.
What Gets Saved
- SNN — Membrane potentials, weights, eligibility traces, thresholds, neuromodulator levels, population definitions
- World Model — Internal SNN weights, prediction history
- GWT — Module activations, broadcast history, winning module
- Emotions — Valence/arousal state, emotional history
- Body Schema — Forward model weights, joint confidences
- Memory — All stored episodes with sensor/motor/reward sequences
- Drives — Drive levels (survival, exploration, comfort, social)
- Metacognition — PE history, fitness history, consciousness level
- Synaptogenesis — Complete concept graph + experience buffer
- Astrocytes — Calcium levels per cluster
- Dream Engine — Replay buffer
- Skills — Active/frozen skills with EWC protection weights
Separate from Checkpoint
Brain state (.pt) stores accumulated cognitive state. Checkpoint stores training-loop state (step counter, CPG phases, competence gate). Both are needed for full resume.
API Reference
save_brain(brain, snn, path, metadata=None)
Serialize complete CognitiveBrain + SNN state to .pt file.
load_brain(brain, snn, path) → dict
Restore all module states. Brain and SNN must be pre-initialized. Returns metadata dict.
brain_info(path) → dict
Quick metadata read without loading full state. Returns version, step_count, consciousness_level, module presence flags, n_synapses, n_episodes, n_concepts.