memento（memento 搜索）

Memento

Usage: /memento [--project PATH] [--user PATH]

Analyze this conversation to extract actionable memory updates for CLAUDE.md and SKILL.md files.

Step 0: Resolve Paths

Default paths:

• Project Memory: ./CLAUDE.md
• User Memory: ~/.claude/CLAUDE.md

Override via arguments:

• --project PATH - custom project memory location
• --user PATH - custom user memory location

Override via config (persistent): Check ~/.claude/CLAUDE.md for a ## Memento Config section:

## Memento Config
- PROJECT_PATH: ./docs/CLAUDE.md
- USER_PATH: ~/custom/CLAUDE.md

Priority: Arguments > Config > Defaults

Store resolved paths as $PROJECT_PATH and $USER_PATH for use in subsequent steps.

Step 1: Analyze the Session

Review the conversation for:

• Mistakes I made - wrong assumptions, incorrect commands, things that failed
• Corrections the user made - "no", "actually", "I meant", "not that"
• Preferences expressed - how the user likes things done (explicit or implicit)
• Project conventions discovered - file structure, tooling, patterns in THIS codebase
• Reusable learnings - insights that apply across all projects
• Improvements for Skills used

Step 2: Categorize Suggestions

What counts as ACTIONABLE:

• Changes how Claude behaves in future sessions
• Specific enough to apply ("use pnpm not npm" vs "check package manager")
• Would have prevented a mistake made in this session

What is NOT actionable (skip these):

• Observations or reflections ("this was a good session")
• Meta-advice ("run retro after meaningful sessions")
• Vague principles ("be more careful")
• Things already known or obvious

Where suggestions go:

**Project Skills/Agents (.claude/skills/*/SKILL.md, .claude/agents/*.md)** - PREFER when:

• Learning is about how a specific skill or agent should behave
• Correction relates to a skill that was used in this session
• Convention improves a specific workflow (biographing, citation, etc.)

Project Memory ($PROJECT_PATH) - Use when:

• Convention is project-wide but not skill-specific
• Architecture or file structure pattern
• General project guidance that doesn't fit a skill

User Memory ($USER_PATH) - Use when:

• Preference applies across ALL projects
• It's about how the user likes to work generally
• It's a Claude Code tip that's universally useful

Step 3: Check for Duplicates

Before suggesting, read:

• Both CLAUDE.md files
• Any skill files (.claude/skills/*/SKILL.md) relevant to the session
• Any agent files (.claude/agents/*.md) that were used

Never suggest something already present in these files.

Step 4: Present Suggestions (or say "nothing meaningful")

If you have ZERO actionable suggestions, just say:

"No actionable suggestions from this session. The conversation was straightforward or learnings were already captured."

If you DO have suggestions, use AskUserQuestion with multiSelect.

IMPORTANT: Show ALL relevant categories in the same AskUserQuestion call.

The questions array can include (only include categories with suggestions):

1. Skills/Agents question (if any skill or agent improvements)
2. Project Memory question (if any project-wide suggestions)
3. User Memory question (if any universal suggestions)

Example structure:

{
  "questions": [
    {
      "header": "Skills/Agents",
      "question": "Which improvements should I add to skill/agent files?",
      "multiSelect": true,
      "options": [
        {"label": "biographing: [suggestion]", "description": "..."},
        {"label": "gardener: [suggestion]", "description": "..."}
      ]
    },
    {
      "header": "Project Memory",
      "question": "Which insights should I add to $PROJECT_PATH?",
      "multiSelect": true,
      "options": [...]
    },
    {
      "header": "User Memory",
      "question": "Which insights should I add to $USER_PATH?",
      "multiSelect": true,
      "options": [...]
    }
  ]
}

For Skills/Agents, prefix each option label with the target file (e.g., "biographing:...").

Rules:

• Max 4 suggestions per category
• Each suggestion must be 1-2 lines, specific, actionable
• User can select none, some, or all from each category

Step 5: Apply Selected Suggestions

For each selected suggestion:

1. Read the target file (CLAUDE.md or skill/agent file)
2. Find appropriate section (create if needed)
3. Append the suggestion using Edit tool

For skill/agent files: Add to an existing relevant section, or create a new section if appropriate (e.g., "## Constraints", "## Tips", "## Common Mistakes").

Output Format

✅ Added to .claude/skills/biographing/SKILL.md:
- [suggestion]

✅ Added to $PROJECT_PATH:
- [suggestion]

✅ Added to $USER_PATH:
- [suggestion]

Or if nothing was added:

No changes made.