Knowledge

Surfaces: web app (Knowledge) · CLI (kcap skills sync, kcap curate apply, kcap memories) · your agent (memory and knowledge tools)

The web app is at https://<your-workspace>.kcap.ai for a cloud workspace; on the local edition the same pages are served from your machine. The desktop app opens it from Help and support → Open dashboard, and kcap local open opens the local one from a terminal

Knowledge is where the team decides what agents should learn. Three kinds of thing live here: facts that evaluations retained, memories people and agents saved on purpose, and docs that group related facts into guidance you can approve as a skill. Choose a scope first; most curation happens at repository scope

What you can do

Review what evaluations learned. Open Knowledge → Docs at repository scope and select Agent if the Agent/Human selector is shown. Each topic has a state, a member count and a pending draft or an approved version. Open one and read three things: the Pending draft, the Serving version it would replace, and the Members, the facts behind it with their source sessions. Edit the draft to remove duplication and say when it applies; Save edit keeps your wording. Docs are the routine route; you do not need to promote every fact one by one

Approve guidance as a skill and install it. Under Approve as, choose Skill, leave Injection off unless the guidance must be present at every session start, and Approve. Then, in each repository checkout where you use agents:

kcap skills sync --dry-run
kcap skills sync

Sync installs the approved versions as kcap-<slug> directories under the repository’s .claude/skills, .agents/skills, .kiro/skills and .gemini/skills, kept out of git through .git/info/exclude, and removes skills the server no longer serves. A skill restricted to one vendor lands only in that vendor’s tree

If you would rather not run the sync by hand, turn on the automatic refresh:

kcap config set skills.auto_sync true

From then on a Claude Code session start refreshes the checkout’s skills in the background, at most once every six hours, so a newly approved skill reaches you within a session or two without another command

Write a rule into CLAUDE.md. For a short rule that belongs in version control, promote it in Knowledge → Queue with Write to CLAUDE.md / AGENTS.md, then in the repository:

kcap curate apply            # shows the diff, asks, writes a managed block
kcap curate apply --dry-run

The block sits between <!-- kcap:curated:start --> and <!-- kcap:curated:end --> and is regenerated on each run; everything outside it is yours. Both files are updated when both exist; AGENTS.md is created when neither does

Skill or CLAUDE.md? Choose CLAUDE.md for a rule every session must follow and that fits in a line or two: it is committed with the repository, so every teammate and every agent that reads the file gets it, whether or not they have kcap. Choose a skill for guidance that is longer, or that only matters for a particular kind of task: the agent loads it when the task calls for it instead of reading it at every start, each checkout picks it up with a sync rather than a commit, and it can be limited to one vendor. A rule that would make CLAUDE.md long is a skill

Find, correct or retire a lesson or memory. Knowledge → Facts lists facts and memories with search and filters for state, category, audience and weight. Open a row to see its source sessions, sightings by repository, where it applies and its curation history. Actions: Promote into guidelines, Dismiss, Undo curation, Mute for me, and Delete or Restore with moderation permission. Memories have Archive and Unarchive instead. A fact can be moved between repository, project and organisation scope, and given a different audience; moving wider may need approval

Save and use memory. Ask your agent to remember something, saying where it applies and who should see it; see What your agent can do. To bring in the memory files Claude Code already keeps:

kcap memories import --dry-run
kcap memories import --audience team --team platform

Decide pending approvals. Knowledge → Queue → Approvals collects decisions that need someone with authority at the wider scope: a fact being promoted from a repository to a project, a nomination that two facts are twins, a request to make something an organisation memory

Decide what new sessions are told. At repository scope, Knowledge → Settings chooses between Promoted-only, which injects curator-approved facts, and Promoted + top-weighted uncurated, which also admits the strongest unreviewed ones. Whichever you choose, only a few documents and facts fit in the session-start briefing; How much it costs gives the limits

Help me get unstuck

Skill is greyed out on a doc. The doc is restricted to particular session kinds or flow roles. It can only be injected; a skill cannot enforce those restrictions

I approved a doc and the agent does not have it. Run kcap skills sync in that checkout. Each worktree needs its own sync

The draft turns one incident into a rule. Edit it, or Dismiss the draft and let the members return to the queue. Revoke an approved doc and choose whether its members stay suppressed or become eligible again

A memory is in the wrong place. Ask your agent to move it to a project or the organisation, or to change its audience. Memories are edited through the agent; the web app archives them

The queue is enormous. Start with Docs instead. Grouped guidance is the point; the queue is for individual corrections and the Regressions lane, which shows patterns that kept growing after promotion