🧠 YecoAI Mini-LLM Cognitive Layer

Ultra-lightweight cognitive protection for Large Language Models

Anti-loop • Amnesia detection • Semantic stability

Developed by www.yecoai.com

The YecoAI Mini-LLM Cognitive Layer is a lightweight, deterministic guard module designed to sit on top of any LLM (GPT-4, Claude, Llama, etc.).

It doesn't replace the model; instead, it acts as a real-time monitoring system that observes, evaluates, and stabilizes the model's output before it reaches the end user or an autonomous agent.

• Infinite Loops: Blocks obsessive repetition of tokens or phrases.
• Context Loss (Amnesia): Detects when the model "forgets" initial instructions or collapses semantically.
• Semantic Degradation: Identifies "word salad" in very long conversations.
• Agent Instability: Protects autonomous workflows from erroneous recursive behaviors.

• 🔁 Multi-level Loop Detector Analyzes structural patterns, n-grams, and semantic repetitions with a precision (F1) of 0.90.
• 🧠 Amnesia Detection Monitors contextual continuity across conversation turns.
• 🧯 Semantic Stability Guard Prevents meaning collapse and nonsensical text output.
• ⚡ Performance Edge Average RAM usage of only 38.85 MB. Ideal for deployment on edge devices and embedded systems.

pip install yecoai-cognitive-layer

This example shows how to use the layer as a "Validator" for a standard LLM response.

from yecoai_cognitive_layer import FeatureEngine, CognitiveModel

# 1. Setup the guards
engine = FeatureEngine()
model = CognitiveModel.load_from_json("weights.json")

def get_safe_llm_response(prompt):
    # Simulate an LLM call (e.g., OpenAI, Anthropic, or Local Llama)
    llm_output = call_your_llm_api(prompt) 
    
    # 2. Cognitive Validation
    vector, features = engine.extract_features(llm_output)
    prediction, scores = model.predict(vector, features)
    
    # 3. Decision Logic
    if prediction == "Loop":
        # If the LLM starts repeating itself, we trigger a retry or a fallback
        return "⚠️ [System Blocked a Loop] Please rephrase your request."
    
    if prediction == "Amnesia" or features['semantic_coherence'] < 0.25:
        # If the response is nonsensical or context is lost
        return "🧠 [Context Loss Detected] I'm having trouble following. Let's restart."

    return llm_output

# Usage
print(get_safe_llm_response("Write a long story about..."))

For autonomous agents, you can use the layer to detect when the agent is "stuck" in a reasoning loop before it consumes too many tokens.

agent_history = []

while agent_running:
    action = agent.think()
    
    # Analyze the agent's thought process
    _, features = engine.extract_features(action)
    
    if features['repetition_score'] > 0.7 or features['struct_loop_flag'] > 0.5:
        print("🚨 Agent Loop Detected! Injecting 'Break Loop' instruction.")
        agent.inject_system_message("You are repeating yourself. Stop and try a different approach.")
        continue
        
    agent.execute(action)

graph TD
    A[LLM Output] --> B[Feature Engine]
    B --> C{Cognitive Model}
    C -->|Normal| D[Validated Output]
    C -->|Loop| E[Block/Regenerate]
    C -->|Amnesia| F[Context Reset]

This project is developed and maintained by YecoAI. Attribution is REQUIRED for any usage, including:

• Modified versions
• Commercial products and SaaS platforms
• Research publications

You must retain this README, the LICENSE file, and the NOTICE file.

Distributed under the Apache License 2.0. ✔ Commercial use allowed ✔ Modifications allowed ✔ Redistribution allowed

YecoAI builds next-generation cognitive systems focused on AI stability and safety.

Website: www.yecoai.com | Discord: Join Community