Playbook OpenClaw

Keep OpenClaw instructions and memory inspectable

Write a narrow workspace instruction, check what memory persists and test recall without treating the workspace as a security boundary.

Read this guide as Markdown

Requirements and responsibilities

Illustrative situation

Illustrative scenario: a small support team wants the assistant to use a fixed public product name in summaries across new chats.

Choose the approach

Dashboard and application procedure

Follow these steps yourself, or use the scoped AI handoff below for supported hosting operations.

Step 1 of 5

Specify a testable instruction

Where
Operations note
Permissions
Authorized OpenClaw owner or application administrator
Inputs
Exact product name and allowed wording
Action
Write one clear rule, such as using a given public product name in summaries, and define a prompt that would reveal whether it was followed.
Expected result
The instruction has a single measurable effect.
Verify
Have an owner approve the wording and scope.
If it fails
If it includes secrets or private data, remove them before any agent write.

Sources: Managed OpenClaw Hosting · OpenClaw agent workspace

Step 2 of 5

Find the actual workspace surface

Where
Deployed OpenClaw app or authorized operator handoff
Permissions
Authorized OpenClaw owner or application administrator
Inputs
Current workspace path or editor and backup owner
Action
If the Control UI is enabled, open Settings → Agents → Files and locate AGENTS.md for the selected agent. Otherwise use a separately authorized operator path; do not assume xCloud has a file editor.
Expected result
The owner knows the real location and access method.
Verify
Read existing instructions before editing and preserve a copy.
If it fails
If the hosted surface is unavailable, record the gap and stop rather than inventing a file path.

Sources: How to Deploy OpenClaw with xCloud · OpenClaw agent workspace

Step 3 of 5

Add the narrow rule

Where
OpenClaw workspace context
Permissions
Authorized OpenClaw owner or application administrator
Inputs
Approved rule and existing file copy
Action
Edit the selected agent’s AGENTS.md operating instructions with the narrow rule, without overwriting existing text; keep credential and channel settings out of the workspace note.
Expected result
The revised instruction is visible in the workspace.
Verify
Compare the diff with the saved copy and confirm only the intended line changed.
If it fails
If the edit replaces other instructions, restore the saved copy and reapply a minimal change.

Sources: Managed OpenClaw Hosting · OpenClaw agent workspace · OpenClaw secrets, storage and logs

Step 4 of 5

Test persistence in a new session

Where
Approved private conversation
Permissions
Authorized OpenClaw owner or application administrator
Inputs
Two prompts about the same public product
Action
Start a new conversation and ask for a short summary that should use the product name. Compare with a control prompt that does not mention the rule.
Expected result
The expected wording appears without pasting the rule into the new prompt.
Verify
Record the observed answer and workspace version.
If it fails
If the agent ignores it, inspect context loading and file placement before adding more memory.

Sources: Managed OpenClaw Hosting · OpenClaw memory architecture · OpenClaw agent workspace

Step 5 of 5

Assign review and removal

Where
Workspace change log
Permissions
Authorized OpenClaw owner or application administrator
Inputs
Rule owner and next review date
Action
Document why the rule exists, how to remove it, and a follow-up test after updates. Check that a delete or correction is reflected in later chats.
Expected result
The instruction remains inspectable and reversible.
Verify
Have a second operator find the rule and repeat the new-session test.
If it fails
If stale wording persists, inspect other memory and session context rather than repeatedly duplicating the rule.

Sources: Managed OpenClaw Hosting · OpenClaw memory architecture · OpenClaw agent workspace

Maintenance

Recovery decisions

  • If the rule is ignored or stale wording appears, compare AGENTS.md with the saved copy and inspect other memory/session context. Revert only the mistaken line; do not restore the entire server for a wording error.

    OpenClaw agent workspace · OpenClaw memory architecture

AI handoff

Use the named xCloud dashboard for hosted controls and verify application tools in the deployed OpenClaw version before use. An xCloud hosting MCP write operation for these actions is not documented.

Supported scope

  • Configure an OpenClaw application behavior app · manual

    Upstream OpenClaw documentation describes application features. xCloud does not document every app configuration surface or pre-enabled tool. Verify the feature and permissions in the deployed build before editing.

    Checkpoint: Use a harmless positive and negative test; preserve current configuration and stop if the feature is unavailable.

    How to Deploy OpenClaw with xCloud · OpenClaw security and access control

Copyable agent brief

Manual checkpoints

  • The app owner edits Settings → Agents → Files → AGENTS.md in an enabled Control UI, or uses an authorized operator path.
  • A second operator tests a fresh chat and verifies the exact instruction is removable.
Feature coverage

Sources

Continue

OpenClaw hosting details