ANT//Agentic Native Task

You Ask. ANT Acts. Memory Persists-Evidence Remains

Comment

💡 Inspiration

Modern AI agents suffer from what we call the "50 First Dates" syndrome—they wake up every session with complete amnesia. While context windows are getting larger, they still lack true continuity and temporal awareness. We realized that memory is not just about storing files; it's about the passage of time and autonomous context retrieval. This inspired us to build MindBy, a sovereign memory satellite within the CluwMind ecosystem that gives AI true continuity.

## ⚙️ What it does MindBy acts as a persistent Agentic Memory Satellite. It transforms stateless LLMs into stateful entities that remember past interactions and understand the passage of time.

• Temporal Awareness: MindBy dynamically injects real-time clock data, allowing the AI to know exactly when you are talking to it (e.g., "Good morning Ard, it's been 12 hours since we last discussed the architecture"). • Autonomous Memory Operations: Through the Model Context Protocol (MCP), external AI agents can autonomously trigger /store and /recall commands. • Semantic Retrieval: When an AI needs context, MindBy searches its distributed memory banks and returns the most semantically relevant past interactions.

## 🏗️ How we built it

We built MindBy with a highly decoupled, modern tech stack designed for agentic workflows:

• Core Engine & Environment: TypeScript and Node.js. Remarkably, the entire development and execution environment was orchestrated on a mobile device using Termux, proving the lightweight yet powerful nature of our architecture. • Database & Vector Search: We leveraged CockroachDB Serverless as our primary persistence layer. Using the native PostgreSQL driver (pg), we utilized CockroachDB's robust VECTOR indexing to store 768-dimensional embeddings. We use cosine distance (<=>) queries to execute lightning-fast semantic searches. • AI & Embeddings: Local embedding models via Ollama (nomic- embed-text) and integration pipelines with AWS Bedrock. • Vibe Coding & AI Collaboration: A massive portion of this project was orchestrated using "Vibe Coding" methodologies. I collaborated heavily with an array of AI models—including Gemini, Claude, Ollama, and Qwen—to rapidly prototype, debug, and architect the entire system. This human-AI synergy was crucial in bringing MindBy to life so quickly. • Interoperability: We implemented the official @modelcontextprotocol/sdk (MCP), exposing our CockroachDB memory functions as executable tools via Server-Sent Events (SSE).

## 🚧 Challenges we ran into

  1. The Latency Illusion: When the AI was querying the CockroachDB vector index and analyzing cognitive tiering, the terminal appeared "frozen." We solved this by building a real-time SSE stream for STATUS: tokens, giving users a transparent view of the AI's internal thought process and database queries.
  2. Temporal Bias: AIs naturally assume every prompt is happening "right now." We had to engineer a dynamic system prompt injection that forces the AI to acknowledge time gaps between sessions.
  3. Integrating MCP: Mapping raw string queries from the AI into precise Vector embeddings and executing them safely against CockroachDB required careful schema design and error handling.

## 🏆 Accomplishments that we're proud of

• Successfully integrating CockroachDB Vector Indexing directly into an autonomous Agentic Loop. • Achieving True Temporal Awareness, making the AI feel like a genuine companion that remembers when you go to sleep and when you wake up. • Building a transparent CLI and Web UI that renders the AI's processes and database operations in real-time.

## 📚 What we learned

We learned that memory is continuity. Coupling a distributed, resilient database like CockroachDB with the Model Context Protocol (MCP) creates a paradigm shift. We no longer have to stuff everything into the prompt context window; the AI can simply "query its own brain" whenever it needs to.

## 🚀 What's next for CluwMind & MindBy

• Cross-Satellite Synchronization: Allowing multiple disparate AI agents (e.g., a coding agent and a scheduling agent) to read and write to the same shared CockroachDB memory pool. • Memory Decay & Consolidation: Implementing background workers that summarize older memories to optimize vector search performance.

Built With

Share this project:

Updates

Submission history