Knowledge bases

A knowledge base is a private store of two kinds of content: documents you add (files or pasted text) and saved conversations — useful bits an agent writes back so a future one can recall them. Both are indexed for semantic retrieval, so an agent can find the passage that means what you asked, not only one that repeats your words.

A knowledge base isn't tied to a single agent. You can attach one to any agent kind, to a Department, to an organization, or bind it as your own personal store that follows you across the agents you use.

When a knowledge base is attached to an agent, the agent gains a search_knowledge tool. Calling it returns the best-matching passages together with a citation for each — the document heading or section trail the passage came from — so an answer can point back to its source. An attachment can also run in auto-inject mode, where relevant passages are prepended before the call without the agent having to ask for them.

An agent can also save the useful part of a conversation into a knowledge base you own, via a save_to_knowledge tool, so later runs can retrieve it — cross-agent memory that is yours. A save only ever lands in a knowledge base you have permission to write; it can never write into a creator's read-only knowledge base.

You create a knowledge base you own, add content to it, and attach it to an agent — from the web, or the equivalent tools in an MCP client:

1. 1
   Create it. Give the knowledge base a name and pick an embedding model — the model that turns text into the vectors used for semantic search. You can choose text-embedding-3-small (the cheaper default) or text-embedding-3-large (higher quality). The choice is pinned at creation.
2. 2
   Bring an embedding key. Ingesting content uses your own OpenAI key from your encrypted vault, so there is no platform charge for embeddings. Until a key is bound, the knowledge base exists but documents won't ingest — nothing is spent on an unconfigured store.
3. 3
   Add content. Add documents by uploading a file (PDF, Markdown, and similar) or by pasting text. Each source is parsed, split into structure-aware chunks, and embedded. Duplicate content is de-duplicated by hash, so adding the same thing twice is a no-op.
4. 4
   Attach it. Attach the knowledge base to an agent or department, or bind it as your personal store. Pick how it's used: as a search tool, as auto-inject, or both. Attaching is not part of the agent manifest, so updating a knowledge base never requires re-publishing the agent.

A knowledge base is scoped — it belongs to a creator, an organization, a department, or a single person's personal store — and that scope is a hard wall, not a hint. A query only ever searches the knowledge bases the caller is actually entitled to: the ones attached to the agent they're running, plus their own personal store. A knowledge base you can't reach is never searched and never surfaces — a scope miss simply returns nothing, with no signal that anything exists.

• Your own memory is always yours. A personal knowledge base is scoped hard to you — a query can never surface another person's personal store.
• A retrieval cap per knowledge base. Because a search embeds the query with the knowledge base owner's key, there's a per-user cap on how much any one caller can retrieve — so a runaway agent can't drain an owner's embedding key.
• The output is scanned. Every retrieved passage passes the same mandatory secret-leak scan the rest of the platform uses before it reaches the model, so a stray secret in a shared store isn't handed back.

A knowledge base attached to a published agent is subject to the same content policy as the agent itself. Because a knowledge base lives outside the agent's manifest, its content can be updated without re-publishing the agent — so it is reviewable on an ongoing basis, not only at submission. If a knowledge base attached to a published agent is found to violate the content policy, it can be suspended: a suspended knowledge base is immediately excluded from retrieval on every agent that uses it, so no agent answers from it until the issue is resolved.

You stay in control of what a knowledge base holds. You can list and delete individual documents, review and delete the conversations an agent saved into a store you own, and delete a whole knowledge base — which also cleans up the underlying stored files, not just the database rows.

Knowledge bases are part of your account data too: your personal knowledge bases and their saved conversations are included in a data export and are removed when you erase your account. A conversation you saved into someone else's (or an organization's) knowledge base keeps the content the owner accumulated, but your authorship link to it is scrubbed.

• Departments — compose agents into a team; a department can share a knowledge base across its members.
• Security model — why declarative agents ship no code, and how the runtime enforces boundaries.
• Creator guide — bring an agent to the marketplace and give it knowledge to draw on.