Knowledge
Attach files and data sources to an agent. They're chunked, vectorized, and retrieved semantically at runtime — so the agent pulls the passages that matter instead of carrying the whole corpus in its context window.
Connect a source once and keep it in sync — no export step, no stale copy to maintain.
Confluence
Google Drive
GitHub
GitLab
Jira
Notion
Zendesk
File Upload
PDFs, docs, exports
Website Crawl
Any public URL
More connectors arriving — if your agent can name a source, the goal is for it to retrieve from it.
A knowledge base is attached to the agent and queried during the run. The model sees the relevant chunks for the question it was actually asked. Context windows stay small, cost stays predictable, and adding a hundred more documents doesn't degrade every prompt.
Upload PDFs and DOCX directly, or connect Confluence, GitHub, GitLab, Jira, Notion, Zendesk, Google Drive, and web URLs. The source stays connected, so the agent's knowledge tracks the system of record rather than a snapshot someone exported last quarter.
Retrieval answers from documents. Memory holds what the agent learned — per-agent persistent notes with titles, bodies, and tags, written and searched through tools during a run. Use knowledge for your corpus and memory for facts about a user, project, or account.
How it fits together
A knowledge base can be attached to an agent directly or bundled into a skill so it loads only when that expertise activates. Collection Pipelines can build an index by mapping an agent across a list-style source, one isolated run per record.
Attach a knowledge base to an agent or scope it to a single skill
Build indexes in batch with Collection Pipelines — each item gets its own trace and cost
Retrieval calls appear in the execution trace like any other tool call
The rest of the build surface
Every capability is included on every plan — see pricing