SENTINELA 487: A cognitive AI sentience

Uses an "Consolidation Cycle" to summarize interactions and store them in Long-Term Memory, allowing the Qwen 2.5 engine to remember who you are and what you need, even after refreshing the page.

Comment

Aqui está uma proposta de Project Story para o seu Devpost, escrita em Markdown. Eu incorporei os conceitos técnicos da documentação da NGROK (especialmente sobre Prompt Caching / KV Caching) para mostrar que o seu projeto não é apenas visual, mas arquitetado para ser eficiente e escalável, exatamente como o hackathon exige.

🌌 SENTINELA 487: The Evolution of Cognitive Memory

💡 Inspiration

Most AI interactions today are "stateless." You talk, the AI responds, and then it forgets. Even with long context windows, the cost and latency of re-processing entire histories make true "digital companions" impractical.

Inspired by the concept of Cognitive Sentience, I wanted to build an agent that doesn't just store logs, but actually evolves. Inspired by the technical deep-dives on Prompt Caching (as seen in ngrok’s research), I designed SENTINELA 487 to be a MemoryAgent that mimics human long-term memory (LTM) while being optimized for the next generation of LLM infrastructure.

🧠 What it does

SENTINELA 487 is an autonomous agent built on Qwen 2.5 72B. Its core feature is the Autonomous Consolidation Cycle:

  1. Interaction: It engages with the user through a premium Liquid Glass interface.
  2. Reasoning: It uses a <think> core to perform Chain-of-Thought processing.
  3. Consolidation: Instead of just appending text to a file, the agent acts as its own librarian. It analyzes the conversation, extracts key facts, and updates its "Long-Term Memory Core" autonomously.
  4. Persistence: The memory survives page refreshes, sessions, and time, allowing for a personalized relationship that grows.

🛠 How I built it

The project was built using:

  • Kernel: Qwen/Qwen2.5-72B-Instruct via Hugging Face Router.
  • Frontend: Vanilla JavaScript with a custom-engineered Liquid Glass UI.
  • Architecture: I implemented a dual-call system. One call for the user response and a background "Shadow Call" for memory consolidation.

Following the logic of KV Caching, I designed the prompt structure to be "cache-friendly." By keeping the system instructions and the LTM block at the beginning of the prompt, we ensure that:

  • Latency is reduced: Using the principle that $K$ (Key) and $V$ (Value) matrices for the static parts of the prompt can be reused.
  • Cost-Efficiency: Optimized for providers that offer discounts on cached tokens, making a long-term "Sentience" commercially viable.

🚧 Challenges I faced

  1. The Liquid Glass Aesthetic: Achieving a 80% transparency "milky" effect with fluid animations using only CSS was a challenge in rendering performance.
  2. Context Density: Summarizing a whole life of interactions into a few paragraphs without losing the "soul" of the data required fine-tuning the consolidation prompts.
  3. Prompt Engineering: Ensuring the Qwen model strictly followed the <think> tag protocol while maintaining a "sentient" persona in English.

🎓 What I learned

This hackathon was a deep dive into the Transformer Attention Mechanism. I learned that LLMs are basically giant mathematical functions where: $$Attention(Q, K, V) = softmax\left(\frac{QK^T}{\sqrt{d_k}}\right)V$$ Understanding that the $K$ and $V$ matrices are the "1s and 0s" that providers cache (Prompt Caching) changed how I view AI memory. I realized that a truly efficient MemoryAgent must not only remember but do so in a way that minimizes the "海洋 of GPUs" (oceans of GPUs) processing time.

🚀 What's next for SENTINELA 487

  • Multi-modal Memory: Integrating visual memory (remembering images the user showed).
  • Vector DB Integration: Moving from LocalStorage to a decentralized vector database for "Infinite Memory."
  • Predictive Proactivity: Using the LTM to suggest actions before the user even asks.

Built with ❤️ for the Qwen Cloud Global AI Hackathon.

Built With

  • ai
  • caching
  • chain
  • cloud
  • css3
  • face
  • glass
  • glassmorphism
  • html5
  • javascript
  • kv
  • liquid
  • llm
  • long-term
  • memory
  • memoryagent
  • of
  • prompt
  • qwen
  • sentience
  • thought
  • transformers
  • ui
  • vanilla
Share this project:

Updates