OP here. I built Orthanc because I was frustrated with the latency of current vector DBs for agentic workflows.

It’s a persistent memory layer designed specifically for any ai agent (not just RAG). It captures the graph of interactions so agents actually 'remember' context over time, not just retrieve chunks.

The Stack:

Sub-200ms retrieval (beating standard vector/Pinecone benchmarks).

Hybrid Graph + Vector approach for context retention.

Open Source SDK / Protocol.

I’m open sourcing the retrieval protocol/SDK today so you can see how the memory graph is structured. I’m currently hosting the backend to manage the graph complexity, but the client is fully open.

Roast my code or let me know what you think.