"Beyond Prompts: A Governance-First, Stateful Persona Substrate for LLMs."
Persona Engine can be used as a persistent cognitive interface between users and computation.
中文文档 | Integration Guide | Architecture | Technical Reports
Persona Engine never decides WHAT to do. It only constrains HOW decisions are formed.
We do not simulate "consciousness" or "autonomy". We provide the Cognitive Runtime that ensures behavioral consistency, auditability, and distinct personality traits across sessions and models.
In the traditional AI stack, Model = Personality. This binds digital identity to a specific vendor's weights (e.g., GPT-4 vs Claude). Persona Engine redefines this architectural relationship:
Model = Compute / Reasoning / Token Factory (Infrastructure)
Persona = Genotype / Memory / Affect Parameters (Identity Layer)
Persona Engine is the Operating System for Digital Identity. By decoupling identity from compute, we create a Portable Persona State that persists across model upgrades, reboots, and platform migrations.
The system is built on the GECCE Kernel, an event-driven micro-kernel ensuring 100% traceability.
| Layer | Component | Function |
|---|---|---|
| L0 | Orchestrator | Lifecycle Management, Persistence, Integrity Checks. |
| L1 | Core FSM | Affective State Matrix (PAD), Intimacy Levels, Governance Barriers. |
| L2 | Genome | Digital Genotype (Probabilistic Loci), Fluid Stance Vectors (R.W.C). |
| L3 | Expression | Seeded Sampling, Prompt Injection, Style Warping. |
| L4 | Memory | Structured Logging, Snapshot Management, Affect-Biased Retrieval. |
A continuous vector model (Rigor, Warmth, Chaos) replaces static "system prompts". The engine dynamically interpolates parameters to adjust the cognitive posture without breaking character.
Utilizing the PAD (Pleasure-Arousal-Dominance) model to quantify "mood" as a verifiable mathematical vector. This allows for:
- Predictable Decay: Emotional spikes normalize over time via homeostasis algorithms.
- Real-time Telemetry: Observation of internal state variables via the dashboard interface.
- Genotype Snapshots: Export the exact state of a persona (including current mood and memory pointers) to a JSON file.
- Audit Logging: Every parameter shift is logged to an append-only journal for compliance review.
Persona Engine is designed to pipe structured cognitive constraints directly into any LLM environment.
# 1. Generate the Persona Context
PROMPT=$(python3 src/persona_cli.py "Explain cybernetics")
# 2. Pipe into your LLM (using tools like 'llm' or 'ollama')
llm -s "$PROMPT" "Explain cybernetics"Output System Prompt (Generated by Engine):
[ROLE]
You are an intelligent AI assistant governed by a dynamic persona engine.
[MISSION]
- Balance abstract theory with practical examples. Be professional and clear.
[POLICIES]
- Confidently stand your ground.
[STYLE]
- Maintain a serious, professional tone. No jokes.
- Balanced and objective.
[OUTPUT_FORMAT]
- Show your work concisely when necessary for correctness.
[OPTIONAL_FLAVOR]
- Occasionally mention interests related to cybernetics.
Persona Engine is designed as passive middleware. It does not execute actions; it computes the context required for your LLM or Agent to act consistently.
graph LR
User["User Input"] -->|Raw Text| PE["🧠 Persona Engine"]
subgraph "Persona Middleware"
PE -->|Computes| Stance["R.W.C Stance"]
PE -->|Retrieves| Mem["Memory Context"]
PE -->|Updates| Mood["Affect (PAD)"]
end
Stance & Mem & Mood -->|"Context Object"| Agent["🤖 Your LLM / Agent"]
Agent -->|"Final Response"| User
Workflow:
- Input: You feed raw user text into the Engine.
- Process: The Engine updates its internal state (mood, memories) but does not generate text.
- Output: It returns a structured Context Object containing the System Prompt, Style Guidelines, and Memory fragments.
- Execute: YOUR system (Agent/LLM) uses this Context to generate the final response.
The following output was generated by
process_interaction("I need to calculate and debug this error").
You are operating under the following cognitive constraints:
[STANCE CONTROL]
- Rigor: High (0.9) — Prioritize factual accuracy.
- Warmth: Low (0.2) — Minimal emotional expression.
- Chaos: Low (0.1) — High determinism.
[AFFECTIVE STATE (PAD)]
- Pleasure: 0.1 (Neutral)
- Arousal: 0.6 (Alert)
- Dominance: 0.8 (Directive)
[BEHAVIORAL GUIDELINES]
- Maintain internal consistency with the current stance.
- Do not speculate beyond provided information.
Persona Engine does not generate responses. It generates the constraints under which responses are formed.
Note: Persona Engine acts as a passive middleware and never executes user intent directly. It generates the context for the LLM to execute.
# 1. Clone and Install
git clone https://github.com/BBQ4ever/Persona-Engine.git
cd Persona-Engine
pip install -r requirements.txt
npm install --prefix dashboard
# 2. Run the Engine & Dashboard (Monitoring UI)
npm run dev --prefix dashboard
# (In a separate terminal)
python3 src/main_demo.pyfrom src.persona_engine import PersonaEngine
# 1. Initialize with a specific Genotype Snapshot
engine = PersonaEngine(snapshot="src/l2_genome/presets/base_persona_v1.json")
# 2. Process an Interaction (Stimulus)
# The engine calculates the new Stance and Affect vectors based on input intent.
# Output is a 'Context Object', ensuring the Engine remains a passive middleware.
context = engine.process_interaction(
user_input="I need technical assistance with this error.",
session_id="user_123"
)
# 3. Inject into Model (Execution)
# The Context Object provides the precise System Prompt parameters.
print(context['system_prompt'])
# Output: "Adopt a High-Rigor [0.9] stance. Focus on factual analysis..."
# 4. Telemetry
print(context['affect_state'])
# Output: {'p': 0.1, 'a': 0.6, 'd': 0.8} (High dominance, high arousal for problem solving)The system has undergone 10 phases of iterative development to validate stability, performance, and architectural decoupling. We are currently implementing Phase 11 (Self-Correction) to introduce governed parameter adjustments based on historical logs.
- Ethics: See ETHICS.md for our approach to safe personality simulation and governance layers.
- License: MIT License.
“Personality is no longer a collection of adjectives, but a computable, observable stream of probability.”