Mastra 发布 Knowledge 自托管服务

Over the last few months we've transformed the way we work at Mastra. When we need to know something that happened, we don't log into a system. We ask one of our agents in Slack: what were our users' responses to last week's new feature? What new objections are surfacing in our sales pipeline?

These are the kinds of questions we've always wanted to know the answers to, and finding the answers usually requires querying multiple systems. With our Slack agents, the answers aren't just in one person's head, they're in everyone's feed.

Last month, we took the next step. Many of these agents were powered by separate knowledge graphs. Then with our software factory, we built a company-wide knowledge graph. All of a sudden, our coding agents had context from the decisions we made about why we were building certain features.

We saw a noticeable improvement in the quality of code shipped. Instead of building things in a naive way, our agents started to build them the way that we'd want.

Building a company brain is hard

Agents are getting easier to build. But constructing a company-wide context layer is hard. Some things we've learned:

• Freshness is difficult and requires versioning. The purpose of an organization is to update important information rapidly, so "checked recently" does not mean "correct now." If you don't include automatic updates, you introduce a base error rate.
• Feature-based organization is needed for complete answers. Topics span systems used by different teams. Even a simple feature might span several code repositories. You can think of a knowledge graph as a local cache of your company's information, and mapping the relationships between systems and topics is difficult.
• Decision records are important. A single feature's record reached 1,054 lines with 19 dated change headings. The decision record preserves the reasoning so future contributors don't unintentionally repeat paths we've already considered.

What Knowledge does

Knowledge is a standalone service that holds company and team context and gives it to an agent when the agent needs it. You connect the systems where your org already writes, like Slack and Monday, and Knowledge keeps itself current.

It sits next to Mastra memory, which handles per-user and per-thread context. Knowledge handles the company and the team. You can run separate knowledge bases per function, like coding and support, and set permissions at the individual, team and org level.

Bring Knowledge to your company

We're starting to work with organizations that want to build their own knowledge bases. All sorts of orgs: a couple are >$50B companies, some are 50-person startups, and there's everything in between. One example: an early design partner feeds questions their support agent couldn't answer back into their company brain, so the agent gets it right the next time.

Today we're opening it up to eight more design partners, first come, first served. If you're building internal agents and thinking about how to expand into a company brain, let's talk.