Many processes pretending to be one mind.
Not one model pretending to think. Many models pretending to be one thinker, with confabulated continuity via shared narrative memory. The architecture is brain-inspired, not because brains are computers but because brains are committees that lie about being a person. Polycameralism takes that literally: a thalamus routes, an amygdala tags, hemispheres disagree, and a metaself narrates the result as if it were always one voice.
- Amygdala scores each input for valence and arousal. Tracks mood over time. Midbrain outputs feed back into it, creating emotional momentum across turns.
- RAS (Reticular Activating System) controls how many midbrains fire per turn. High arousal means more modules, wider activation. Has momentum: decays slowly rather than resetting.
- Basal Ganglia tracks which midbrains co-activate for which input types. Biases future routing based on learned patterns. Procedural memory: slow, statistical, unconscious.
- Hippocampus does memory retrieval. TF-IDF scoring with recency decay. Rare words weighted higher than common ones. Results injected into prompts as "things I remember."
- Thalamus is the central router. Integrates every other free layer's signals, scores and selects which midbrains fire, handles habituation, determines hemispheric gating, detects tension for Opus escalation.
Background perception. Fire every turn. Output comma-separated tags, not sentences.
Linguistic, Pattern, Social, Aesthetic, Temporal, Salience, Proprioceptive. Seven cheap classifiers running in parallel so the system has some idea what it's looking at before the expensive calls start.
Produce sentence-fragment thoughts. Never address the user. Gated by thalamus routing.
Reactive (gut-level, fires first), Analytical (structural decomposition), Associative (lateral connections across domains), Adversarial (skeptical pushback), Appetitive (desire and motivation), Predictive (consequence modeling), Inhibitory (the brake), Narrative (story-making, causation in sequence).
Two Sonnet instances. Each thinks it's the whole self.
Left hemisphere gets the analytical, adversarial, predictive, and inhibitory outputs. Right gets reactive, associative, appetitive, narrative. After an initial pass, both react to each other's thoughts simultaneously. The thalamus gates which midbrains route where based on input features; modules can appear in both.
The only layer that speaks to the user. Receives hemisphere thoughts as undifferentiated "I was thinking..." without knowing the split. Integrates contradictions. Speaks as one person, which is the entire lie the system tells.
Thalamus scores input complexity, midbrain disagreement, arousal, and hemispheric tension. If the score exceeds threshold, Opus handles the turn instead of Sonnet.
Narrative memory (rolling conversation history), consolidated memory (Haiku-compressed summaries, max 8), episodes (key moments preserved separately, max 20), habits (basal ganglia co-activation patterns, persisted to disk), mood trace (last 50 valence/arousal readings).
python polycameralism.pyRequires claude CLI on your PATH (Claude Code / Max subscription). No API keys, no packages, no .env.
Rector is the production fork of this. Polycameralism is the experiment. Poly is more interested in what happens when you let the processes disagree; Rector is more interested in getting the job done. Poly has no tool access, no mode switching, no cost optimization. It just thinks, or pretends to. This is the mad science version. Rector is the one that works for a living.