A bi-temporal content library for AI agents

Teams standing up AI agents rarely think carefully about how the content those agents draw from is managed. Most RAG/agent setups dump documents into a vector store with no lifecycle metadata: no record of when a document actually took effect, no clean way to roll back a bad content push, no way to stage a change before it reaches production. The result is agents that confidently cite policies that haven't taken effect yet, that were superseded weeks ago, or that were quietly wrong and can't be reverted in one step.

System prompts, instructions, and other materials an agent depends on to run at all are usually hard-coded or pasted into a config somewhere, with no versioning, no effective-dating, and no rollback story -- a bad prompt push is just as capable of silently breaking agent behavior as a bad policy push, and today it's harder to fix.

Get a sandbox Read the docs →

Free sandbox, seeded with sample content. No card required.

Every version is effective-dated -- rollback is publishing an older one, live on its next call.

Rollback is one action, not a redeploy

Every version of a document, policy, or system prompt is append-only and effective-dated. Publishing a rollback is the same create → mark ready → publish flow as any other change -- pick an older version's content, set an effective date, publish. No pointer to repoint, no manual data fix, no redeploy.

Effective-dated: Set when a version takes effect and stage a draft ahead of time. The same effective_at/as_at pins also let you point a connection at a past moment -- audit what was live then, or re-run your agent against it to see what would have happened.
One version of truth per agent: Every MCP endpoint resolves against a fixed library and the current moment -- an agent can't be prompted into picking a different version of the truth.
Documents, prompts, and tool definitions -- one model: Reference documents, system prompts, and even a tool's own name/description/argument text all get the same versioning, effective-dating, and rollback treatment -- not just the parts that look like documents.