I build custom AI agents for businesses, and one of the things I've noticed is that everyone wants to use AI to do real, critical work, but it's hard to trust it because sometimes agents can go a little off the rails. No one wants their agent deleting all their customer data or screwing up their books. So they tend to limit its use to searching, summarizing, and as a writing assistant. This really limits the value you can get from using AI.
The problem is that most ways we store data, from notes & spreadsheets to databases, were designed for humans who work at and make mistakes at human speed, or for deterministic software systems. AI agents work non-deterministically at machine speed, which is a new challenge.
That's why I made Jaybase. Jaybase is an append-only fact store for agentic workflows. It's based on a Merkle Directed Acyclic Graph, which is the same technology behind git and the blockchain, so every write is auditable, replayable, and reversible. If an agent goes off the rails and corrupts a bunch of data, it's easy to recover and easy to audit what went wrong. It also doesn't have a fixed schema, so as you add agents or your agents get more skills, it's natural to add more kinds of data.
Happy to dive into any questions in the comments and would love the community’s feedback. Thanks!