I've been running my own knowledge base for AI agents for months now. Markdown files, mostly. It works great right up until your data stops being clean and small.
Real data isn't clean or small. It's scattered across Notion, Linear, Slack, call transcripts, and half of it contradicts the other half. Dumping all of it into a bigger context window and hoping the model finds the needle doesn't actually work - a bigger haystack isn't an easier search. The retrieval is the hard part, and that's not what the model is built to do.
So it was good timing to see
@DeytaHQ come out of stealth this week with Khora. It's an open-source library for giving agents durable memory: it pulls in messy multi-source data, builds a knowledge graph plus vector store under the hood, and routes each query to whatever retrieval actually fits. Hybrid GraphRAG, you bring your own database.
What makes me want to test it: it's solving the exact problem I keep hand-rolling at OneBalance and in my own setup. Connecting facts a human would naturally connect in their head, so the agent reasons over real knowledge instead of a pile of text.
It also helps that I know the people behind it. I worked with
@MiljanTekic at Tenderly, where he was a co-founder, so I've seen how he builds dev tooling up close. His co-founder Igor Bogicevic is a serial entrepreneur - he co-founded Seven Bridges (genomics data) and Orgnostic (people analytics, later acquired by Culture Amp). The through-line across all three companies is making messy, heterogeneous data usable. That's a good track record to bet on this problem.
I'll put it head to head with my markdown memory and report back. If you're building long-horizon agents, it's worth a look.