sivanhavkin/Entelgia: Unified AI core for persistent agents, internal conflict, and moral self-regulation through dialogue.

Entelgia is a psychologically-inspired, multi-agent AI architecture designed to explore persistent identity, emotional regulation, internal conflict, and moral self-regulation through dialogue.

This repository presents Entelgia not as a chatbot, but as a consciousness-inspired system — one that remembers, reflects, struggles, and evolves over time.

Two primary agents engage in continuous, persistent dialogue driven by a shared memory database, allowing emergent internal tension and moral reasoning to arise naturally — rather than executing pre-defined rules.

Complete rewrite with a production-ready architecture

This version represents a full architectural overhaul focused on robustness, performance, and long-term extensibility.

• Multi-agent dialogue system (Socrates · Athena · Fixy)

• Persistent memory

  • Short-term memory (JSON)
  • Long-term memory (SQLite)

• Psychological drives

  • Id / Ego / Superego dynamics

• Emotion tracking & importance scoring

• Dream cycles & memory promotion

• LRU caching (≈ 75% hit-rate improvement)

• REST API interface (FastAPI)

• Unit testing (9 tests, pytest)

• 10-minute auto-timeout

• PII redaction & privacy protection

• Resilient error handling with exponential backoff

• Full structured logging

• ~50% reduction in LLM calls via caching
• ~70% reduction in token usage via compression
• 2–3× faster response times

• ~1,860 lines of production-ready code

• 25+ classes with full type hints

• 50+ documented functions

• Modular core system:

  • Memory
  • Emotion
  • Language
  • Conscious
  • Behavior
  • Observer

Entelgia stores memory locally on your machine using SQLite and JSON files.

entelgia_data/
├── entelgia_memory.sqlite    # Unified long-term memory database
├── stm_*.json                # Per-agent short-term memory
├── entelgia_log.csv          # Session logs and interaction history
├── entelgia_graph.gexf       # Memory graph exports (optional)
└── versions/                 # Version history

To delete all stored memory and reset the system:

1. Stop the program

2. Delete the SQLite database:

   rm entelgia_data/entelgia_memory.sqlite

3. Delete per-agent short-term memory files:

   rm entelgia_data/stm_*.json

4. (Optional) Clear logs and graphs:

   rm entelgia_data/entelgia_log.csv
   rm entelgia_data/entelgia_graph.gexf
   rm -rf entelgia_data/versions/

5. Run the system again — files will be recreated automatically.

Entelgia uses terms such as consciousness, emotion, conflict, and self-regulation as architectural metaphors, not claims of biological or phenomenological consciousness.

These concepts describe internal system dynamics such as:

• Memory prioritization and revisitation
• Competing objectives within dialogue
• Observer-based correction loops (meta-cognitive feedback)

The goal is not to simulate a mind, but to explore how complex internal structure and moral tension can emerge in autonomous AI systems through design.

• ✅ Unified AI core implemented as a single runnable Python file (Entelgia_production_meta.py)
• ✅ Persistent agents with evolving internal state
• ✅ Emotion- and conflict-driven dialogue (not prompt-only)
• ✅ Dialogue continuity across sessions via shared memory
• ✅ Meta-cognitive monitoring and corrective feedback loops

Reflective, questioning, and internally conflicted. Drives inquiry through doubt and self-examination. Serves as the primary agent for exploration and dialectical reasoning.

Integrative and adaptive. Synthesizes emotion, memory, and reasoning, providing coherence and emotional context to Socrates’ inquiry.

An architectural role designed to detect loops, errors, and blind spots, injecting corrective perspective shifts to prevent stagnation or logical fallacies.

• A research-oriented architecture inspired by psychology, philosophy, and cognitive science

• A system modeling identity continuity, not stateless interaction

• A platform for experimenting with:

  • Emotional regulation
  • Moral conflict and resolution
  • Self-reflection and meta-cognition
  • Meaning construction over time

• Not a chatbot toy
• Not prompt-only roleplay
• Not safety-through-censorship
• Not a replacement for human judgment or ethics review

Central Premise: True regulation emerges from internal conflict and reflection, not external constraints.

Instead of relying on hard-coded safety barriers, Entelgia emphasizes:

• Moral reasoning through dialogue
• Emotional consequence tracking
• Responsibility and repair mechanisms
• Learning through error rather than suppression

Consciousness as Process: Consciousness is treated as a process, not a binary state. The system explores how reflective dialogue, memory continuity, and internal tension create emergent cognitive properties.

Entelgia is organized around six interacting cores:

1. Conscious Core — self-awareness, reflection, narrative construction
2. Memory Core — unified persistent SQLite memory with STM/LTM stratification
3. Emotion Core — dominant emotion detection, intensity, and regulation
4. Language Core — dialogue-driven cognition and adaptive phrasing
5. Behavior Core — goal-oriented intentional responses and consequence tracking
6. Observer Core (Fixy) — meta-level monitoring and corrective intervention

[Session Start - Memory Loaded]

SOCRATES: "Athena, we revisited our discussion about intention yesterday.
          I notice I'm still uncertain: does responsibility require
          the ability to have chosen otherwise?"

ATHENA:   "Your uncertainty is not a flaw, Socrates. But remember—we also
          explored how emotional commitment shapes choice. Perhaps the
          question isn't about abstract possibility, but about what we
          genuinely care about."

[Emotion tracking: Socrates = Contemplative (0.7), Athena = Integrative (0.8)]
[Memory update: "responsibility-intention-link" promoted to long-term]

SOCRATES: "That's different from what I concluded before. Let me reconsider..."

Entelgia explores ethical behavior through dialogue-based internal tension, not enforced safety constraints.

• Ethical dynamics emerge implicitly through agent interaction
• Conflicting moral frameworks are expressed through dialogue
• Errors and contradictions trigger reflection and memory promotion

• Researchers exploring consciousness-inspired AI architectures
• Developers interested in persistent multi-agent dialogue systems
• Philosophers & psychologists studying computational models of self
• Contributors curious about AI systems that do more than respond

Requirements-

• Python 3.10+
• Ollama with a local LLM (e.g., phi3, mistral, neural-chat)
• 8GB+ RAM recommended (16GB+ for larger models)
• pip install requests colorama fastapi uvicorn pytest networkx
• Entelgia will automatically attempt to install missing Python dependencies at runtime for convenience.

Entelgia runs entirely on a local LLM for privacy, control, and reproducibility. You must install Ollama before running the system.

Download Ollama for your operating system:

👉 https://ollama.com

Supported platforms:

• macOS
• Linux
• Windows (WSL recommended)

After installing Ollama, pull at least one supported model:

Recommended models:

• phi3 (3.8B) – Fast, low memory, ideal for testing
• mistral (7B) – Balanced reasoning and performance
• neural-chat (7B) – Strong conversational coherence
• openchat (7B) – Fast and stable dialogue

💡 On systems with 8GB RAM, prefer phi3. Larger models may be slow or unstable.

Run a quick test:

If you see a response, Ollama is installed and working correctly.

git clone https://github.com/sivanhavkin/Entelgia.git
cd Entelgia
ollama serve
python Entelgia_production_meta.py

Released under the Entelgia License (Ethical MIT Variant with Attribution Clause).

The original creator does not endorse or take responsibility for uses that contradict the ethical intent of the system or cause harm to living beings.

Conceived and developed by Sivan Havkin.

Contributions are welcome. Please open an issue or discussion before submitting major changes.

• Status: Production / Research Hybrid
• Version: v1.0
• Last Updated: February 7, 2026