Persistent Agent Memory on CockroachDB

Inspiration

AI agents often lose context between sessions. I wanted to build an agent that truly remembers conversations and can retrieve them across sessions, using a production-grade database as its memory layer.

What it does

This project demonstrates an agentic memory system that stores conversation history and embeddings in CockroachDB. It retrieves memory through the Managed MCP Server and a serverless AWS Lambda function.

How we built it

-CockroachDB Cloud stores chat history as JSONB and supports vector search. -The Managed MCP Server lets AI agents query the database directly. -AWS Lambda retrieves stored memory and exposes it via a public Function URL. -Distributed Vector Indexing enables semantic search on embeddings.

Challenges we ran into

-Getting vector indexing to work smoothly with CockroachDB. -Making the Lambda function connect securely to CockroachDB. -Ensuring the MCP server correctly returned stored messages.

Accomplishments that we're proud of

-Successfully storing and retrieving persistent memory. -Demonstrating real agent access via MCP. -Deploying a working serverless function on AWS.

What we learned

How CockroachDB unifies transactional data and vector search in one system, and how MCP simplifies agent-to-database interactions.

What's next for Persistent Agent Memory on CockroachDB

Adding real-time embeddings for better semantic recall, integrating with Bedrock agents, and scaling to multi-region deployments.

Built With

• ai
• aws-lambda
• cockroach
• langchain
• mcp
• postgresql
• python
• vector-search