Skip to main content
The Knowledge Base is available for orgs on the skills + persistent-memory architecture.
The Knowledge Base is a page in the workspace sidebar — alongside Data Model, Skills and System Prompts — that holds reference documents the agent must read before it can interpret your business correctly: definitions, business rules, policies, company background, and known caveats about the data. It answers a different need than System Prompts. A system prompt is a compact set of rules the agent applies on every message. The Knowledge Base is a library of reference material the agent searches on demand — long-form documents that would swamp a system prompt but are invaluable when a question touches them. If your refund policy is two pages long, it belongs in the Knowledge Base, not the system prompt.
A workspace's Knowledge Base page showing uploaded markdown files with tags

​
Adding files

Files can be added two ways from the Knowledge Base page: Create inline — write a new document directly in the built-in markdown editor. Best for definitions and rules you’re authoring for the agent in the first place. Upload — drop in existing .md, .txt, or .docx files. Uploads are converted to Markdown so that every file is editable in the same editor, diffable between versions, and readable by agents without a parser. The original file is not stored.
Upload zone for dropping markdown, text, or Word documents into the Knowledge Base
Organize files with tags so related documents stay grouped as the library grows.

​
What belongs in the Knowledge Base

  • Definitions — the canonical meaning of terms users will use loosely (“gross margin”, “active account”, “churn”)
  • Business rules and policies — refund windows, SLA tiers, pricing rules the agent should cite rather than guess
  • Company background — how the business is structured, what entities relate to what
  • Data caveats — “the legacy_orders table stops at March 2024”, “EMEA revenue excludes partner channels”
The test is: would the agent need to read this to interpret a term correctly? If yes, it belongs here. If it’s a short rule that must always apply without being looked up, it belongs in the system prompt instead.

​
Versioning

The Knowledge Base is version-aware, just like the rest of the workspace. There is one working copy per workspace, and writes land on the draft version. Files are checkpointed when you publish a workspace version, so every published version carries the exact set of documents the agent had at that point — and edits never silently alter a version that’s live in production. See Workspaces for how draft and production versions work.

​
How the agent uses it

At chat time, the agent doesn’t preload your Knowledge Base. It reaches in with two tools:
  • search_context_files — finds files relevant to the user’s question
  • read_context_file — reads a specific file’s contents in full
So the cost of a large library is low: documents only enter the agent’s context when retrieval says they’re relevant. This is also why descriptive file names and good tagging matter — they improve what retrieval surfaces.

​
Next steps

Data Model & Schema

The foundation the Knowledge Base sits in — tables, columns, descriptions, and versioning.

Skills

Package procedures and conventions the agent loads on demand.

System Prompts

Short, always-on rules that apply to every message.

Golden Assets

Example questions with trusted SQL and charts, retrieved by similarity.