CogArch 2.0 Cognitive Architecture
Brain-like thinking
Neurova uses the CogArch 2.0 cognitive architecture, modelled on how the human brain processes information.
The four cognitive centres (mirroring brain regions)
| Brain region | Concept | Responsibility |
|---|---|---|
| Cerebral cortex | Cognition | Observing and understanding, memory recall, logical reasoning, deciding actions, self-reflection |
| Cerebellum | Planning and coordination | Decomposing intent, generating tasks, orchestrating execution, evaluating results, recovering from errors |
| Brainstem | Action output | Calling tools, running workflows, scheduling resources, monitoring execution |
| Spinal cord | Information pathway | Dispatching events, connecting modules, attaching external channels |
The full cognitive loop
input → observe → recall → reason → decide → orchestrate → execute → reflect → consolidate → learn and evolveThe stages in detail
- Observe — perceive external input, understand context and intent
- Recall — retrieve relevant memories, experience and emotional associations
- Reason — reason logically from internal state, memory and emotion
- Act — turn decisions into real behaviour through cerebellar orchestration and brainstem execution
- Reflect and consolidate — evaluate the outcome, extract experience, strengthen memory, correct the model
The shared core
Each agent owns an independent brain (its database) and office (its working directory), but all agents share one Shared Core: the cerebellum (Plan Orchestrator), the brainstem (Execution Engine) and the spinal cord (Infrastructure).
Design principle: cognitive autonomy with resource reuse — agents reason and decide separately while sharing orchestration and execution.
The emotion engine
Neurova introduces an emotion engine built on psychological classification theory, forming a complete system of four tiers and 17 emotions.
The four tiers
- Basic emotions (5): joy, sadness, anger, fear, surprise
- Composite emotions (4): affection, envy, compassion, disgust
- Higher emotions (4): shame, guilt, pride, sense of responsibility
- Special emotions (4): love, hate, hope, despair
Key properties
- Emotion propagation rules — 17 rules governing how emotions influence one another
- Emotion-weighted decisions — different emotions carry different weight in a decision
- Integration with memory — emotional state shifts memory temperature
Recursive Self-Improvement (RSI)
Neurova treats "improving its own capability" as a first-class concern, implemented through RSI (recursive ratchet pruning):
- Ratchet pruning: parameter optimization only ever converges in a better direction
- Convergence analysis: improvement stops automatically once its returns diminish
- Progressive rollout: shadow mode → canary release → full release
- Multiple safety boundaries: parameter bounds + rate-of-change limits + rollback + human oversight