RoachMind — Persistent Agentic Memory RAG Engine

RoachMind indexes research paper embeddings directly inside CockroachDB using native vector search (VECTOR(384) with <-> cosine distance).

On every query, RoachMind:

• Performs real-time distributed vector retrieval against CockroachDB running in aws-ap-south-1.
• Synthesizes relevant context chunks cleanly.
• Automatically writes complete conversation turns, session states, and timestamps into a persistent conversation_memory table.

How we built it

• Database & Memory Layer: CockroachDB Cloud (aws-ap-south-1) with native vector search and psycopg2.
• Agent Architecture: Built in Python with sentence-transformers for embedding generation and rank-bm25 for hybrid context re-ranking.
• AWS Integration: Integrated via boto3 for cloud deployment and LLM synthesis.
• Frontend: Interactive Streamlit web app deployed on Streamlit Community Cloud.

Challenges we ran into

Building a production-grade fallback pipeline that seamlessly switches between cloud LLM inference and local vector synthesis without throwing runtime errors or disrupting user sessions.

Accomplishments that we're proud of

• Zero-downtime persistent memory layer that scales horizontally in the cloud.
• Sub-second distributed vector retrieval directly inside CockroachDB SQL engine.
• Deploying a fully functional, public production build on Streamlit Cloud.

What we learned

How to design fault-tolerant agent architectures using CockroachDB's native vector data types and distributed database mechanics.

What's next for RoachMind

• Multi-region cluster scaling for global agent deployment.
• Expanding memory pruning and summarizing policies for multi-year agent retention.

Built With

• amazon-web-services
• boto3
• cockroachdb
• llm
• postgresql
• python
• rag
• streamlit
• vector-search