mneme

• 

  mneme

• 

  Agent Architecture

• 

  agents

• 

  agents

Inspiration

Memory is the real bottleneck in AI. Developers struggle with context windows, but in fields like healthcare or research, losing context means losing critical insight. Even with a million-token models, complex thinking needs persistent memory, not temporary context.

What it does

MNEME solves the context window problem by giving AI a second brain. It remembers past knowledge, combines it with live research, and produces outputs that improve over time instead of resetting every run.

How I built it

I built a multi-agent system where each agent has a single role. Neo4j stores structured memory, Gemini handles reasoning, and agents coordinate through a clean orchestration layer with real-time UI showing the full flow.

Challenges we ran into

Managing both frontend and backend alone was tough. UI stability issues slowed me down, and coordinating real-time agent updates while keeping the system consistent was harder than expected.

Accomplishments that I am proud of

A fully working system where agents visibly collaborate in real time. The UI shows each step live, making the system feel like it’s thinking instead of just responding.

What I learned

Loose coupling is everything. Treating each agent like a microservice made debugging easier and the system more scalable. Clear responsibility per agent beats one overloaded system every time.

What's next for mneme

I want to take this to production. Right now retrieval uses string matching; next step is integrating semantic retrieval or RAG to improve entity understanding and make memory truly intelligent.

Built With

• fastapi
• google-gemini-api
• google-search-grounding
• neo4j
• next.js
• python
• react
• tailwind
• typescript
• websockets