Guidance and memory

Give every agent the same standing instructions, and control what agents remember about you, a project or the whole workspace.

On this page

Agents read several kinds of standing guidance before they start a task: the workspace’s general instructions, your communication habits, the repo’s own agent files, the agent’s system prompt and anything they have been allowed to remember. You manage the workspace parts in Settings → Guidance & memory.

What an agent reads, in order

  1. General instructions from Settings, given to every agent in every project.
  2. Communication habits, how agents should talk to people.
  3. Repo instructions such as AGENTS.md or CLAUDE.md in the repo the run works in.
  4. The agent’s system prompt (see Agents).
  5. The workflow step’s instructions, with {{references}} filled in.
  6. Memories in scope for this run, after you approved them.

Later, more specific guidance wins when two conflict. A step that says “skip the changelog for this run” beats a general instruction to always write one.

General instructions

General instructions are given to every agent, in every project. Use them for rules that hold across the workspace:

Lead with the decision I need to make.
Say what you haven't verified.
Never merge on a Friday after 16:00 Europe/London.
Use British English in user-facing copy.

The editor shows the number of rules and “applies from the next run”. Runs in progress keep the instructions they started with. Save to see “Instructions saved · new runs use them”.

Keep these short. Each line costs tokens on every run, and long instructions dilute the important ones. Anything specific to one repo belongs in that repo’s AGENTS.md; anything specific to one role belongs in the agent’s system prompt.

Communication habits

Communication habits are small, reusable preferences about how agents write to people: in asks, PR descriptions, Slack posts and run summaries. Examples offered in Settings:

  • Lead with the decision I need to make
  • Numbered priorities
  • Say what you haven’t verified

Add your own as short phrases. Habits are per workspace and apply to every agent’s messages, not to code.

Repo instructions

Agents read the repo’s own instruction files when a run starts: AGENTS.md, CLAUDE.md and the files they point to, such as CONTRIBUTING.md. Duun doesn’t change these files; they live with the code and are reviewed like code. If an agent proposes a change to them, it does so in a pull request.

Memory

Agents notice things while they work: that a test suite needs a database, that you prefer small PRs, that a service has a quirk. Memory lets them keep what they learn, with your approval.

Memory scope

Pick one scope for the workspace:

ScopeWhat it means
PersonalYour agent remembers you across every conversation. Nothing carries over to teammates.
ProjectYour agent learns each project and shares what it knows inside that project. Memory doesn’t cross projects.
Shared workspaceAnyone in the workspace can read and write what agents remember.

Changing the scope shows “Memory scope updated” and applies to new memories and new runs.

Proposed memories

Agents propose memories as they work. You approve each one before it’s used. Proposals show up in What agents remember with the run that proposed them, and you can approve, edit or dismiss each one. Nothing an agent proposes is used until a person approves it.

Good memories are facts that save time later:

  • “The payments test suite needs docker compose up ledger-db first.”
  • “Dana reviews all changes under services/ledger/.”
  • “Staging deploys freeze during finance reconciliation on Mondays.”

Forget

Select memories and choose Forget to delete them. Forgetting is permanent and takes effect from the next run. To clear everything in the current view, use Forget N memories? from the list’s menu.

Memories are workspace data. They are covered by the workspace’s export and deletion settings in Settings → Storage & privacy, and they are never used to train models. See Security and data handling.

    Type to search the docs, help center and blog.