Memory

Memory Skill

CRITICAL: Use This Skill

You MUST use memory during task execution. Files are your only persistent storage.

At task start:    Read .memory/MEMORY.md
During execution: Append to .memory/tasks/<task_id>.md
At task end:      Update .memory/MEMORY.md with reusable learnings

Use filesystem.write_text_file or filesystem.append_text_file MCP tools to write memory.

Purpose

Record observations, findings, and learnings during task execution to:

1. Avoid repeating mistakes - Remember what didn't work
2. Track progress - Know what you've already tried
3. Preserve context - Don't lose important findings across compaction
4. Build knowledge - Accumulate learnings for future tasks

Memory Types

Short-Term Memory (Task-Specific)

Location: .memory/tasks/<task_id>.md

Record task-specific observations as you work:

• Elements found on screen
• Actions attempted and their results
• Errors encountered
• Posts/content discovered
• Navigation paths tried

When to write:

• After each significant action
• When discovering important information
• When encountering errors or obstacles
• Before moving to next step

Long-Term Memory (Cross-Task)

Location: .memory/MEMORY.md

Record durable knowledge that applies to future tasks:

• App UI patterns (where buttons are located)
• Common error solutions
• Successful strategies
• User preferences
• Platform-specific quirks

When to write:

• Discovering a reusable pattern
• Finding a workaround for a common issue
• Learning something about app behavior
• At task completion (lessons learned)

Memory Format

Short-Term (Task) Memory

# Task: <task_id>
**Started**: <timestamp>
**Goal**: <task description>

## Progress Log

### <timestamp> - <action_type>
**Action**: <what you did>
**Result**: <what happened>
**Observation**: <what you learned>

### <timestamp> - Finding
**Source**: <where you found it>
**Content**: <key information>
**Relevance**: <why it matters>

## Summary
<brief summary of findings when task completes>

Long-Term Memory

# MobileAgent Memory

## App Knowledge

### Threads (com.instagram.barcelona)
- Search: Tap magnifying glass top-right, not bottom nav
- Feed: Swipe up to scroll, tap post to open
- Back: Android back button or swipe from left

### WeChat (com.tencent.mm)
- Official Accounts: Me > Settings > Official Accounts
- Article list: Tap account name > View history

## Strategies

### Research Tasks
1. Always scroll 3+ screens before concluding
2. Open at least 5 posts for thorough research
3. Check comments for additional context

## Error Solutions

### Unicode Input Fails
- Ensure DeviceKit is installed
- Use mobile_type_keys with submit=true

## User Preferences
- Language: zh-TW
- Output format: Structured with sources

Usage Instructions

At Task Start

1. Read long-term memory for relevant knowledge
2. Create task memory file with initial template

MCP filesystem commands:

// Step 1: Read long-term memory
filesystem.read_text_file({ path: ".memory/MEMORY.md" })

// Step 2: Create task memory (use session ID as task ID)
filesystem.write_text_file({
  path: ".memory/tasks/<session_id>.md",
  content: "# Task: <description>\n\n## Observations\n\n"
})

During Task Execution

After each significant action, append to task memory:

MCP command:

filesystem.append_text_file({
  path: ".memory/tasks/<session_id>.md",
  content: "\n### Screen Analysis\n- Found search at (543, 150) top-right\n- 15 clickable elements\n"
})

Example memory entry:

### Screen Analysis
- Found search at (543, 150) top-right, NOT bottom nav
- 15 clickable elements including: Home, Search, Profile
- Feed shows 3 posts, need to scroll for more

When Finding Information

Record key findings immediately so you don't lose them:

filesystem.append_text_file({
  path: ".memory/tasks/<session_id>.md",
  content: `
### Content Found
- Source: @clawdbot_user
- Content: "Claude MCP integration is amazing for automation"
- Engagement: 234 likes, 45 comments
- Sentiment: Positive
`
})

At Task Completion

1. Append summary to task memory
2. Update long-term memory with NEW learnings only

MCP commands:

// Append task summary
filesystem.append_text_file({
  path: ".memory/tasks/<session_id>.md",
  content: "\n## Summary\n- Found 8 posts, 6 positive\n- Themes: AI, MCP, productivity\n"
})

// Update MEMORY.md if you discovered something new and reusable
// Read first, then append new section
filesystem.read_text_file({ path: ".memory/MEMORY.md" })
filesystem.append_text_file({
  path: ".memory/MEMORY.md",
  content: "\n### New App: Reddit\n- Search via magnifying glass\n- Posts sorted by Hot/New/Top\n"
})

Before Context Gets Too Long

If you've been working for a while, flush important state to memory:

filesystem.append_text_file({
  path: ".memory/tasks/<session_id>.md",
  content: `
### Progress Checkpoint
- Posts reviewed: 3 of 5 target
- Sentiment so far: 2 positive, 1 neutral
- Still need: Check @user1, @user2 posts
- Next action: Scroll to find more posts
`
})

This ensures you can resume if context is compacted.

Best Practices

1. Write early, write often - Don't wait until the end
2. Be concise - Memory is for reference, not prose
3. Use structured format - Easy to parse later
4. Separate facts from opinions - Quote sources
5. Update long-term memory sparingly - Only durable knowledge

File Paths

.memory/
├── MEMORY.md           # Long-term memory (cross-task)
└── tasks/
    ├── abc12345.md     # Task-specific memory
    ├── def67890.md
    └── ...

Integration with Other Skills

• search-triage: Record search results in task memory
• content-extract: Save extracted content summaries
• comment-scan: Log sentiment analysis findings
• patrol: Track visited posts to avoid duplicates