AEON-Δ: A Cognitive Architecture for Emergent Reasoning

Welcome to the official repository for AEON-Δ, a modular and self-contained cognitive architecture engineered in PyTorch. AEON-Δ simulates high-level reasoning by operating on abstract "thought vectors" (z-vectors) within a latent space, enabling deep reflection, concept modeling, and plan generation.

The system follows a rigorous two-phase training pipeline, designed for stability, abstraction, and dynamic reasoning, with each component engineered for extensibility, robustness, and clarity.

---

🛠️ Core Components & Cognitive Engine (aeon_core.py)

1. AEONDelta (nn.Module)

Central orchestrator class that integrates all reasoning modules into a cohesive flow from perception → deliberation → planning → action.

2. ThoughtEncoder / ThoughtDecoder (nn.Module)

Latent autoencoder translating between text and compressed thought vectors (z). This forms the bedrock of reasoning operations.

3. QualiaExtractor (nn.Module)

Transforms the initial thought vector into a raw perception ψ_0, serving as the entry point into the reflective reasoning cycle.

4. MetaLoopProcessor (nn.Module)

The reflective engine. Iteratively transforms ψ_0 into a stable thought state C* through repeated refinement, simulating recursive internal deliberation.

5. PillarsModule (nn.Module)

Dissects C* into 5 interpretable core axes of meaning ("pillars"): 🔥 Will, ⚔️ Resolve, 🔄 Growth, 🛡️ Union, 🌊 Movement.

6. QuantumSimulator (nn.Module)

Calculates entanglement, coherence, and complexity of internal representations using quantum-inspired metrics.

7. TopologyAnalyzer (nn.Module)

Applies catastrophe theory to detect instabilities and sudden representational shifts — critical for modeling insight, rupture, or contradiction.

8. Action & Planning Modules (nn.Module)

Uses the refined representation to:

• Produce an action vector
• Estimate its safety
• Generate a high-level plan

9. RSSM: Recurrent State-Space Model

Captures thought dynamics by predicting how z_t evolves into z_{t+1} across time.

---

📀 Data & Knowledge Integration (aeon_core.py, ae_train.py)

• Robust Data Loader: Handles JSON, NDJSON, and concatenated formats.
• Tokenizer: Simple character-level tokenizer transforms raw text into tensors.
• Curriculum Pipeline: Separates short/long sequences for progressive learning.
• MemoryManager: Built-in support for mem0 vector memory.
• KnowledgeGraph: Optional Neo4j-backed graph for external knowledge.

---

⚙️ Training Pipeline (ae_train.py)

Phase A: Geometry of Thought (SafeThoughtAETrainer)

• Trains autoencoder (Encoder/Decoder)
• Uses curriculum learning (short → full sequences)
• Dynamic function routing for backward compatibility
• Logically safe and clean loss computation

Phase B: Dynamics of Thought (FixedZDynamicsTrainer)

• Uses trained encoder to convert text into z-vector sequences

• Trains RSSM and core dynamics to predict next z from current z

• Introduces robust loss functions:

  • kl_diag_gaussians: avoids mode collapse
  • cosine_spread_surrogate: promotes diverse, orthogonal thought states

---

🛠️ Engineering Utilities & Robustness

• Structured Logging: Clean JSON logs
• Sanitization Filters: Remove invalid/control characters
• Deduplication: Prevent spam logs
• CLI Interface: Via argparse for clean configuration
• Safety Checks: Assert tensors, NaN/Inf filters, contiguity assertions

---

🚀 Mission

AEON-Δ is built to model not just cognition, but emergent reasoning: how thoughts form, evolve, refine themselves, and lead to action. This is not merely a transformer wrapper — this is a full cognitive simulator, ready to grow.

Pull requests and collaborations welcome.

No bits left behind. — AEON-Δ

∆: No bits left behind. It begins with the choice to be.