Teach CloudThinker

Seven surfaces hold the context you give CloudThinker: workspace memory, Dreaming, Incident Memory, Learnings, knowledge bases, skills, and runbook sources. This page picks the right one for what you’re teaching, and each row links to the page that sets it up.

Which one to use

Surface	What it teaches	Who writes it	Where it applies	Reach for it when
Workspace Memory	Durable facts about your environment, team preferences, decisions, and corrections	CloudThinker, as agents work. Anyone who can chat teaches or corrects it; workspace Admins edit and delete items directly	Workspace-wide. Loads into every conversation	You’re tired of re-explaining your setup, or an agent keeps using a fact that’s wrong
Dreaming	Nothing new. It keeps workspace memory and custom skills short, current, and consistent	CloudThinker, nightly. A workspace Admin turns it on or off and tunes it	Workspace-wide: memory items and custom skills	Memory has filled up with duplicates and stale items — or your team curates by hand and you want it off. On by default
Incident Memory	What each investigation found: symptom, root cause, the fix, and a check that proves the fix	CloudThinker, at the end of an investigation. There’s no form to fill in and nothing to switch on	Resolve only	The same service keeps failing the same way and you want the next investigation to start from the last one
Learnings	Your team’s review standards — as rules distilled from resolved findings, and as the convention files you already keep in the repository	CloudThinker writes Learnings; your team writes the convention files it reads. A workspace Admin sets the schedule and retires a rule	Review, one repository at a time	Review keeps raising something your team fixes by hand, or your standards are already written down in the repository
Knowledge Base	Your documents, as written: runbooks, policies, architecture notes, crawled web pages	You, by uploading files or crawling a site. A workspace Admin creates them and adds documents	Workspace-wide. Every agent can search every knowledge base, and Resolve can use them as evidence	Agents should answer from your document instead of generic advice
Skills	A procedure or set of rules agents follow step by step	Usually you. CloudThinker can also write them, through a custom agent’s Skill Learning setting, `/create-skill`, or Dreaming	Workspace-wide, but each skill runs only on the surfaces ticked under Available in: Chat, Review, Incident, Assessment	The same procedure should run the same way every time
Runbooks — Legacy	Fix procedures you already keep in Confluence, GitHub, or GitLab, or upload as Markdown	You. The procedures stay in your own systems; a workspace Admin adds the source	Resolve only, during root cause analysis	Your fix procedures already live in one of those systems and you don’t want to rewrite them as skills

Runbooks carries a Legacy label in the app. It works as described on its page, and no replacement or removal date has been announced. The other six surfaces are current.

I want to…

I want to	Use
Make agents follow our runbook, every time	Write the runbook as a skill and tick the surfaces it applies to
Use the fix procedures we already keep in Confluence or a repository	Add them as a runbook source, or upload the same files to a knowledge base if you’d rather stay off a Legacy surface
Stop Review repeating a comment we keep dismissing	Find the rule in Review → Learnings and retire it
Teach Review a standard it keeps missing	Write it in a convention file in the repository
Give CloudThinker our architecture doc	Upload it to a knowledge base
Stop re-explaining which region production runs in	Say it once in chat and let workspace memory keep it
Correct a fact an agent got wrong	Say the correction in chat; workspace memory updates the matching item
Reuse what we learned from the last outage	Nothing. Incident Memory already reads the last investigation before it forms a hypothesis
Keep memory from filling up with duplicates	Leave Dreaming on, and add a focus instruction if a run keeps getting something wrong

How they fit together

Workspace memory is the foundation. Incident Memory and Learnings are specialized memory for one module each, written by CloudThinker rather than by you, and neither appears in the workspace memory list. One test separates the two surfaces you write yourself: if agents should follow it step by step, make it a skill; if they should look things up in it, put it in a knowledge base. A skill can point agents at a knowledge base document, so the two work together.

Examples

Teach or correct workspace memory by saying it in any chat. The reply shows a Saved to memory or Updated memory step.

Production runs in eu-west-1 and disaster recovery runs in eu-central-1.
Don't propose rightsizing for the staging cluster. Why: it's sized for load tests.

Name a knowledge base in your request to limit the search to that one.

Using the Payments runbooks knowledge base, how do I fail over the primary Postgres?

Customize

Shape agents with instructions, skills, and connections

Workspace Memory

Browse and correct what CloudThinker remembers

Skills

Turn a procedure into something agents apply on their own

Core concepts

See how agents, approvals, and context fit together

Connections

Customize

Approvals

Workspace settings

Admin Settings

Run anywhere

Artifacts and sharing

Which one to use

I want to…

How they fit together

Examples

Customize

Workspace Memory

Skills

Core concepts

​Which one to use

​I want to…

​How they fit together

​Examples

​Related

Customize

Workspace Memory

Skills

Core concepts

Which one to use

I want to…

How they fit together

Examples

Related