Skip to main content
Knowledge is where agents keep information they learn and search for later. Personal and workspace Knowledge work the same way; only access differs.

Facts versus behavior

Knowledge holds facts. A request like “keep replies concise from now on” changes behavior instead, so it belongs in:
  • Workspace instructions for short rules that apply to every session. These are added to every prompt, so keep them brief.
  • A Skill for a procedure or a preference that applies only in some contexts. Only its description is in the prompt; the agent reads the rest when relevant.
The agent tells you whether a change is active, waiting for review, or not saved.

What Knowledge contains

The agent saves things worth keeping, such as a decision, a requirement, or the cause and fix of an incident. Uploads and screenshots are not saved automatically. With learning enabled, you do not need to say “remember.” Adopted decisions, lasting product feedback and corrections can update existing entries. A tweak requested only for the current task stays in that task. The agent retrieves relevant published Knowledge when later work depends on those decisions. New Slack chats follow the same learning settings. Older Slack chats can retain their earlier explicit-save restriction. A referenced file’s name alone does not give the agent its contents; it must read the file before retaining its details.
  • File: an original document, such as a PDF.
  • Source: a reference document, or evidence behind a finding.
  • Finding: a piece of information, such as a renewal date, linked to its source.
  • Collection: related entries about one customer, product, or system. Collections can nest, and an entry can sit in several.
The Knowledge page has Library, Instructions, and Review tabs. The agent checks existing entries before saving and updates them rather than adding duplicates. Scheduled tasks save Knowledge the same way chats do. New chats with Knowledge enabled can search it directly, without first discovering the search tool. Search starts with five short matches; the agent opens relevant entries for the full text and can ask for more results. Older chats keep their existing tool setup and conversation history. To find new knowledge, open Library → Filter → Added and choose Last 24 hours, Last 7 days, or Last 30 days. The filter uses each entry’s original creation date, not its latest edit. It works with search and the other filters; remove the Added chip to show entries of any age.

Personal and shared Knowledge

A private chat saves to its user’s personal Knowledge and can read shared Knowledge that user has access to. A shared chat saves to the workspace. In the SDK, the chat option memory: "workspace" | "user" | false picks the scope. It defaults to "user" for private chats and "workspace" for shared chats. false turns off saving; the agent can still read Knowledge. Personal Knowledge requires a server-verified user. See Users, tenants & privacy.

Agent learning and review

Settings → Agent learning (also in Knowledge’s ⋯ menu) controls what agents can add or change: Each category can be:
  • Automatic: changes publish immediately.
  • Review first: changes wait in the Review tab. The agent keeps working without asking in chat. You approve, edit, or reject each change, grouped by chat turn or scheduled run.
  • Off: agents make no new changes. Existing Knowledge, Skills, and human edits are unaffected.
Workspace and personal defaults apply unless a chat (+ → Chat settings) or a scheduled task overrides them. These settings control what agents write, not what they read, and they don’t replace tool approvals.

Conversation history and task notes

Conversation history is the session transcript. Task notes help agents coordinate within one session tree. Neither is Knowledge: save lasting facts in Knowledge, procedures in a Skill, and rules in Instructions.