Nexus Memory

Inspiration

AI agents are getting better and better at doing things like researching writing code and checking their work.. When many agents are working on their own a big problem comes up: what happens to their knowledge if an agent stops working crashes or gets replaced?

We thought about it. Came up with the idea of treating AI agents like a group of people working together. A research agent should be able to share what it knows with a coding agent. A validation agent should be able to learn from mistakes.. A new agent should be able to take over without having to start from nothing.

This idea led us to create Nexus Memory. A way for multiple AI agents to share knowledge in a way.

Our main idea is simple:

Agents are temporary. Memory is not.

What it does

Nexus Memory lets many different AI agents work together by using a shared memory that stays the same.

A normal process has three agents:

• Research Agent. Looks into the task and finds information from work.

• Coding Agent. Uses the research and past information to make a solution.

• Validation Agent. Checks the solution for errors, security problems and things that could be better.

Of keeping important details inside an agents short-term memory the system saves the task state agent results, choices, security problems and past experiences in a shared database.

Nexus Memory also uses search so that agents can find past experiences based on what they mean not just on exact words.

The system has a test where an agent can fail and a new agent can take over from where the old one left off.

How we built it

We made Nexus Memory as a system for AI agents, with a website where people can watch how the agents work.

The system has this structure:

User

↓

Task Orchestrator

↓

Research Agent → Coding Agent → Validation Agent

↓

Shared Memory Layer

↓

CockroachDB

Built With

• ai
• amazon-web-services
• api
• cockroachdb
• embeddings
• gemini
• google
• javascript
• mcp
• node.js
• react
• rest
• search
• studio
• typescript
• vector