Cloud Skills

Procedural memory and skill architecture

Cloud Skills provide user-authored procedural memory in XMemo, enabling developers and teams to save standard operating procedures (SOPs), runbooks, checklists, and reusable prompts directly into persistent storage. Once saved, connected agents across supported tools can discover, recall, and interpret them without per-agent installation or manual prompt copying. A stored Cloud Skill is distinct from the installed XMemo Skill client runtime (@xmemo/skill, node scripts/xmemo-skill.mjs), which serves as the agent-side connection bridge.

• Preview capability: Cloud Skills is currently in Preview in the Memory Console (#skills).
• Identity and asset: Each skill has a unique skill_id, human-readable slug, name, description, owner scope (Personal or Team), and asset status (active or archived).
• Immutable revision chain: A skill contains an append-only sequence of SkillRevision snapshots. Each revision carries a revision_id, revision_number, instruction_body, canonical_hash, and optional typed SkillComponent resources.
• Size limits: Root instruction bodies are limited to 64 KiB UTF-8. A revision can contain up to 50 declared components, with each component limited to 128 KiB.
• Component resource types: Components declare explicit semantic types including reference, example, template, document, artifact, and script.

Publication lifecycle and team review

Cloud Skill revisions transition through a structured lifecycle to protect production workflows and support collaborative authoring:

• draft: Initial revision state during authoring and private staging before public activation.
• proposed: Revisions authored within a Team workspace transition to proposed status, awaiting explicit review and activation by a team owner or admin.
• published: The active revision eligible for agent discovery and default Level 1 recall. Only one revision per slug is designated as published at any given time.
• deprecated: Older or superseded revisions can be marked deprecated to signal sunset while preserving historical reproducible recall.
• Personal vs Team publishing: In Personal workspaces, saving with publish intent transitions directly to published. In Team workspaces, proposed revisions prevent unreviewed procedural changes from altering agent workflows.

Progressive recall: Level 0, Level 1, Level 2

Connected agents recall skills on demand using Progressive Recall to prevent saturating LLM context windows with unused documentation:

• Level 0 (Discovery): Query search returns lightweight descriptors (slug, name, description) matching the user task without loading full instructions.
• Level 1 (Root Recall): Recalls the skill root instructions and component manifest for a specified slug, defaulting to the published revision with latest-revision fallback.
• Level 2 (Resource Recall): Recalls a single declared component (such as a checklist template or reference schema), requiring the caller to supply an explicit revision_id to prevent resource tearing.

recall_cloud_skill(query="incident response")  # Level 0: lightweight descriptors
recall_cloud_skill(slug="incident-response-checklist")  # Level 1: root instructions + manifest
recall_cloud_skill(
  slug="incident-response-checklist",
  revision_id="<revision-id-from-level-1>",
  resource="templates/triage-template.md",
)  # Level 2: one pinned resource component

Revision pinning prevents resource tearing

When a skill is updated, XMemo creates the next immutable revision rather than overwriting historical content. Level 1 root recall delivers root instructions alongside an immutable revision_id and component manifest. Level 2 resource requests require the caller to pin that exact revision_id. If an author publishes a new revision while an agent is executing a workflow, the agent pinned read continues to resolve the exact snapshot it started with, eliminating resource tearing.

{
  "kind": "root",
  "data": {
    "skill_id": "skill_inc_triage_01",
    "revision_id": "rev_v1_inc_triage",
    "revision_number": 1,
    "canonical_hash": "c7b14d8e9f2a4c10...",
    "resources": [
      {
        "logical_path": "templates/triage-template.md",
        "type": "template",
        "content_hash": "e3b0c44298fc1c14..."
      }
    ]
  }
}

Import, export, and authoring surfaces

Developers can author Cloud Skills interactively in the web console or import existing skills from standard distribution packages:

• Console authoring: Create skills interactively in Memory Console (#skills) with markdown editor and component management.
• Package import: Import standard SKILL.md files (with YAML frontmatter), multi-file ZIP archives, or remote repositories via URL (ClawHub or GitHub).
• Revision diffing: Compare instruction changes and component modifications across any two revisions with word-level diffing.
• Source Capsule export: Export any skill revision as a lossless Source Capsule ZIP archive containing SKILL.md, component resources, and metadata manifest.

Agent access profiles and sandbox isolation

Authorized agents access Cloud Skills through contractually defined MCP tools in supported client environments:

• Contracted MCP tools: The tools recall_cloud_skill (read-only) and save_cloud_skill (authoring) are contractually exposed in the cursor-plugin and ordinary-full client profiles (config/mcp-documentation-contract.json). They are not exposed in generic-public, chatgpt-public, or claude-code profiles.
• Prompt-executable procedures: Most Cloud Skills are standard operating procedures interpreted directly by agents as prompt guidance without code execution.
• Sandbox execution boundary: Optional backend script execution via execute_cloud_skill is an advanced server-side capability governed by SandboxExecutionEngine (not exposed to any client profile today). When local bubblewrap container isolation is absent, execution fails closed with SANDBOX_UNAVAILABLE.

Availability and lifecycle boundaries

Cloud Skill capabilities are available if procedural Cloud Skills are enabled for your account. Revisions are immutable content-addressed snapshots, and saving changes creates a new revision without overwriting prior records. When candidate skills exist, the console displays a review banner requiring explicit manual promotion.