Workspace Memory

Browse, correct, and import what CloudThinker remembers about your workspace so agents stop asking for the same context twice

Workspace memory is the set of notes CloudThinker keeps between conversations: facts about your environment, team preferences, decisions, and corrections. Every new conversation starts with it loaded, so you explain your setup once and every teammate benefits.

Dreaming is on by default. It runs nightly, merges and deletes memory items and custom skills, and uses workspace credits each night it has conversations to read. Turn it off if you’d rather curate memory by hand.

What CloudThinker remembers

Agents write memory themselves while they work. You don’t have to tag anything, and raw transcripts are never copied into memory.

Memory	What it holds	Example
Memory items	Durable facts, decisions, and corrections, typed Workspace, Feedback, Project, or Reference. Shown in More → Workspace settings → Memory	”Production runs in eu-west-1; disaster recovery runs in eu-central-1”
Connection notes	What an agent learned while working with a specific connection. Not shown in More → Workspace settings → Memory. Ask in chat to see or change them	Which log group holds the payment service logs
People preferences	How each member likes to work: report format, tone, priorities. Not shown in More → Workspace settings → Memory. Ask in chat to see or change yours, for example “Change my report preference to bullet points"	"Keep answers short and lead with the result”

Each memory item has one type. Pick the type by what the item is:

Type	What goes there	Example
Workspace	Conventions, glossary, infrastructure facts, ownership	”The payments team owns the pay-* RDS instances”
Feedback	Corrections to how agents work, fixes that worked, approaches to avoid	”Don’t propose rightsizing for the staging cluster. Why: it’s sized for load tests”
Project	Initiatives, decisions, deadlines	”Migration to Graviton finishes by 2026-12-15”
Reference	Pointers to external systems, docs, or dashboards	”Cost anomalies are triaged in the FinOps Grafana dashboard”

Ongoing work is also tracked as a task: a memory item that holds the task’s status (active, blocked, or done), decisions, blockers, and next steps. Tasks show in More → Workspace settings → Memory like any other item, and each one gets a one-line entry in the memory index (see Limits).

Two products keep their own specialized memory on the same foundation: Incident Memory for Resolve investigations and Learnings for Review. To choose between memory, knowledge bases, skills and runbook sources, see Teach CloudThinker.

Memory is shared by every member of the workspace, not private to you. Agents can also search every member’s past conversations in the workspace, and there’s no way to mark a conversation private. Don’t paste secrets or personal data into chat. Store secrets as credentials instead.

Who can do what

Action	Who
Browse and search memory	Every workspace member
Add memory through chat, Tell memory something…, or Import	Members with permission to chat (not Viewers)
Edit or Delete an item directly	Workspace Admins

The Admin lock covers direct edits only. Any member who can chat can ask CloudThinker to change, delete, or restore an item, and there’s no setting to restrict that short of giving the member the Viewer role. Every save is recorded in the item’s revision history with the member, agent, and conversation behind it. To see who changed an item, ask in chat: “Who changed the on-call handover item, and what did it say before?”

See roles for how roles are assigned.

Browse memory

1. Open the Memory tab

   Go to More → Workspace settings → Memory.

2. Find an item

   Scroll the Memory list, grouped by type and newest first, or type in Search memory. Each row shows the item’s name (for example cost-report-routing), set by the agent that created it, and its one-line summary. Search matches names and summaries, not the full details.

   
3. Open the item

   Click a row to see Last updated, the one-line Summary, and the Details in Markdown.

   

Success state: you can see each fact CloudThinker uses about your workspace and when it last changed. The same page also holds the Dreaming controls and Run history.

Teach or correct memory

The fastest way to change memory is to tell CloudThinker, in any chat or from the Memory tab. Each correction applies to every later conversation in the workspace.

Where	What to do	What happens
Any chat	Say the fact or correction in plain words	CloudThinker updates the matching item, or creates a new one if none matches. The reply shows a Saved to memory or Updated memory step
Tell memory something… (bottom of the Memory tab)	Type a note and send it	A Memory chat opens, saves the note into the right item, and lists what it changed
Ask CloudThinker to change this… (inside an item)	Describe the change	A Memory chat opens, edits that item, and lists what it changed

Forget that we exempt the staging cluster from rightsizing.
Cost reports go to #finops now, not email.
Our on-call handover is Fridays at 16:00 UTC.

If the reply has no Saved to memory or Updated memory step, open More → Workspace settings → Memory and check the item’s Last updated time.

Edit or delete an item

Workspace Admins can change an item directly, without a chat turn. Other members see Edit and Delete locked and use Ask CloudThinker to change this… instead.

1. Open the item

   In More → Workspace settings → Memory, click the item.

2. Edit it

   Click Edit. Change the Summary (one line, up to 200 characters), the Type, or the Details (Markdown), then click Save. The name is set by the agent that created the item and can’t be edited here. To rename an item, ask CloudThinker in chat.

   
3. Or delete it

   Click Delete and confirm Delete this memory item?. CloudThinker forgets the item for every member. A copy stays in the revision history.

Success state: the item shows the new summary and an updated Last updated time, or it disappears from the list.

To get a deleted or badly edited item back, there’s no restore button. Ask in chat instead, naming the item: “Restore the cost-report-routing memory item from its history.” Any member who can chat can do this. History keeps up to 20 revisions per item by default and prunes the oldest when a new change lands, so restore soon. There’s no time-based expiry and no export. On a self-hosted install, an instance admin can change the revision count.

If a secret lands in memory

Import and agents are told to skip secrets, but the agent decides what counts as one, so a secret can still slip through.

1. Delete the item, or ask CloudThinker to remove the secret from it.
2. Rotate the secret in the system it belongs to. Deleting the item doesn’t remove copies already in its revision history, and the app has no control to purge them, so treat the secret as exposed. To ask about removing it from history, email [e-mail] with the item name.
3. Store the new value as a credential.

Import memory from another AI tool

Bring what another assistant already knows about your team, so you don’t start from zero.

1. Open Import

   In More → Workspace settings → Memory, click Import (or the + in the Tell memory something… box). The Import memory to CloudThinker dialog opens.

2. Copy the prompt

   Click Copy and paste the prompt into a chat with your other AI tool. It asks that tool to export everything it knows about your team, infrastructure, projects, and conventions as Markdown.

3. Paste or upload the answer

   Paste the answer into the text box, or drag files into the drop zone. Pasted text is limited to 200,000 characters; upload anything longer as a file. You can add up to 5 files per import in .md, .txt, .json, .csv, or .pdf.

   
4. Add it

   Click Add to memory. A Memory chat opens and CloudThinker splits the content into items, merges it with what already exists, and skips secrets and credentials. A Memory chat (from Import, Tell memory something…, or Ask CloudThinker to change this…) is a regular chat and uses credits like one. See Usage.

Success state: the Memory chat ends with a list of the items it created or changed, and they appear in More → Workspace settings → Memory.

Limits

Item	Value
Summary	One line, up to 200 characters
Details	Up to 20,000 characters of Markdown
What loads into each conversation	The memory index (a one-line entry per item, first 200 lines) and people preferences. Lines past 200 aren’t loaded, but agents can still search those items. Agents open an item’s full details only when the work needs them
Task entries in the index	Up to 20. At 20, agents drop a finished or less relevant task’s entry to make room. The task item itself stays and can still be searched
Items listed in More → Workspace settings → Memory	The 150 most recently updated

Ask about past conversations

Agents can search and read earlier conversations in the same workspace and use them as evidence. Ask “Which chats last week mentioned the payments RDS instance?”, or paste a CloudThinker chat link and ask the agent to continue from it.

Search covers conversations from every member of the workspace, not only yours, so treat workspace chats as visible to the team. It never reaches other workspaces, deleted conversations, or the conversation you’re in. A conversation can’t be marked private or excluded, so deleting it is the only way to keep it out of search. Old instructions found this way are treated as background facts, so your current request always wins.

Prompt suggestions in the composer

When recent conversations left open next steps, the empty chat composer shows up to three suggested prompts, each tagged with the conversation it came from. Suggestions draw on roughly the last week of workspace activity. Typing filters them.

Action	Result
Click a suggestion	The prompt is inserted with a context line that points to the source conversation. Leave that line so the agent can read it
Click x on a suggestion	It never appears for you again. Teammates still see it, and memory doesn’t change

There’s no workspace setting to turn suggestions off. An instance admin (whoever runs a self-hosted CloudThinker install; on CloudThinker cloud, email [e-mail]) can switch them off for the whole instance.

Troubleshooting

The tab says Memory is not ready yet

Workspace memory couldn’t be read at that moment. Your items are safe. Click Retry after a few seconds.

Memory is updating. Showing the last saved copy.

A recent chat or Dreaming run wrote to memory and the list hasn’t caught up. The page refreshes on its own every few seconds until the new version shows.

Saving shows This item changed since you opened it

Dreaming or a chat edited the item while you had it open. Click Reload to see the new version, then make your change again. Your unsaved edits are replaced.

Edit, Delete, or Import is locked

Edit and Delete need the workspace Admin role. Import and Tell memory something… need permission to chat, which Viewers don’t have. Ask a workspace Admin.

An agent keeps using an outdated fact after I corrected it

Another item probably still holds the old fact. Type a keyword from it in Search memory, then edit or delete the conflicting item (or ask CloudThinker to). If it comes back the next morning, open the latest run in Dreaming Run history: the summary lists the items it wrote. Add a focus instruction that names the correct fact.

The list says No memory yet

CloudThinker hasn’t saved anything for this workspace. Chat with it about your environment, send a note through Tell memory something…, or import from another tool.

FAQ

Is memory per user or per workspace?

Per workspace. Every member’s conversations read and write the same memory. Only prompt-suggestion dismissals are personal.

Can I turn memory off, clear it, or export it?

Not per workspace. A workspace can turn off Dreaming, but not memory itself. An instance admin can switch off workspace memory for the whole instance. There’s no bulk clear or export: delete items one by one in More → Workspace settings → Memory, or ask CloudThinker in chat to remove them.

How is memory different from the knowledge base?

You upload documents to the knowledge base and agents search them when needed. Memory is written by agents from what they learn while working, and it loads into every conversation automatically.

Next steps

Dreaming

Control the nightly run that merges and prunes memory

Knowledge Base

Upload runbooks and docs agents can search

Incident Memory

Reuse lessons from past investigations in Resolve

Agents

See how CloudThinker and custom agents use memory