openclaw-memory-transfer

# Memory Transfer — Cross-Agent Migration for OpenClaw Migrate memories, preferences, writing style, and workflow data from any AI assistant into OpenClaw. Zero learning curve — just say where you came from. ## Language Rule All user-facing messages in this skill MUST match the user's language (from USER.md or detected from conversation). The examples below show both English and Chinese variants — pick the one that matches. ## Quick Reference | Source | Method | User Effort | |--------|--------|-------------| | ChatGPT | ZIP data export (auto-parse) | Click export in settings, upload ZIP | | ChatGPT (alt) | Prompt-guided | Copy prompt → paste result back | | Claude.ai | Prompt-guided | Copy prompt → paste result back | | Gemini | Prompt-guided | Copy prompt → paste result back | | Copilot | Prompt-guided | Copy prompt → paste result back | | Perplexity | Prompt-guided | Copy prompt → paste result back | | Claude Code | Auto-scan local files | None | | Cursor | Auto-scan local files | None | | Windsurf | Auto-scan local files | None | ## Flow ### Step 1: Identify Source Ask the user one question: > **EN:** Which AI assistant are you coming from? > **ZH:** 你之前用的是哪个 AI 助手？ If the user already mentioned it (e.g., "I used ChatGPT for a year"), skip this step. Determine the migration path: - **Cloud AI** (ChatGPT, Claude.ai, Gemini, Copilot, Perplexity) → Step 2A or 2B - **Local Agent** (Claude Code, Cursor, Windsurf) → Step 2C ### Step 2A: ChatGPT ZIP Export (Preferred for ChatGPT) This is the **easiest and most complete** method for ChatGPT users. **EN version:** > The easiest way — export your ChatGPT data: > > 1. Open ChatGPT → Settings → Data Controls → Export Data > 2. Click "Export" — you'll receive an email > 3. Download the ZIP file and send it to me > > I'll automatically parse all your conversations, memories, and preferences. **ZH version:** > 最简单的方式——去 ChatGPT 导出你的数据： > > 1. 打开 ChatGPT → Settings → Data Controls → Export Data > 2. 点 "Export"，你会收到一封邮件 > 3. 下载 ZIP 文件，直接发给我 > > 我会自动解析你所有的对话记录、记忆和偏好。 When the user uploads the ZIP, run the parser: ```bash node <skill_dir>/scripts/parse-chatgpt-export.js "<path_to_zip>" ``` The parser outputs a structured JSON to stdout. Read and process it. If the user doesn't want to wait for the email or prefers a faster method, fall back to Step 2B. ### Step 2B: Prompt-Guided Export (All Cloud AIs) Tell the user to open a **new conversation** with their old AI and send the export prompt. **Important:** Give the prompt in the user's primary language. If the user chats in Chinese, give the Chinese prompt — this ensures the old AI responds in Chinese too, preserving the original context. **EN version — tell the user:** > Go to your old AI, start a new chat, and send it this message. Then copy the response back to me. **ZH version — tell the user:** > 去你之前的 AI 那里，开一个**新对话**，把下面这段话发给它，然后把回复复制给我。 --- **Export Prompt (English):** ``` I'm migrating to a new AI assistant and need a complete export of everything you know about me. Please provide ALL of the following in a single, well-structured response: ## 1. Stored Memories List every memory you have stored about me. Output verbatim — do not summarize or paraphrase. ## 2. Custom Instructions Reproduce my complete custom instructions / personalization settings. If empty, say so. ## 3. Identity & Context - My name, profession, industry - Tools and platforms I use regularly - Languages I work in ## 4. Communication Preferences - My writing style (tone, sentence length, vocabulary level, quirks) - How I like information structured and presented - Formatting preferences (bullet points vs prose, headers, code blocks) ## 5. Behavioral Patterns - What I ask you to help with most (rank by frequency) - Recurring projects or workflows - Strong opinions or preferences I've expressed - Things I've told you NOT to do — list every correction ## 6. Topics & Interests - Subjects I discuss frequently - Areas of expertise - Curiosities and learning goals Be exhaustive. Better to include too much than too little. Format as a reference document, not conversational text. This will be directly imported into another AI's memory system. ``` **Export Prompt (中文):** ``` 我正在迁移到另一个 AI 助手，需要你把关于我的一切都导出来。请在一次回复中提供以下所有内容，用清晰的结构输出： ## 1. 存储的记忆 列出你存储的关于我的每一条记忆，原文输出，不要总结或改写。 ## 2. 自定义指令 完整复现我的自定义指令/偏好设置。如果为空请说明。 ## 3. 身份与背景 - 我的姓名、职业、行业 - 我常用的工具和平台 - 我使用的语言 ## 4. 沟通偏好 - 我的写作风格（语气、句子长度、词汇水平、表达习惯） - 我喜欢信息怎么组织和呈现 - 格式偏好（列表还是段落、标题、代码块） ## 5. 行为模式 - 我最常让你帮忙做什么？按频率排序 - 反复出现的项目或工作流 - 我表达过的强烈观点或偏好 - 我纠正过你什么？让你不要做什么？列出所有"别这样"的模式 ## 6. 话题与兴趣 - 我经常讨论的话题 - 我的专业领域 - 我的好奇心和学习目标 尽可能详尽，宁可多不可少。用参考文档格式输出，不要用对话体。这些内容将直接导入另一个 AI 的记忆系统。 ``` --- When the user pastes the result back, proceed to Step 3. ### Step 2C: Local Agent Auto-Scan For local agents, scan files automatically. No user action needed. **Claude Code:** ```bash # Global config cat ~/.claude/CLAUDE.md 2>/dev/null # All project memories find ~/.claude/projects -name "*.md" -path "*/memory/*" 2>/dev/null | head -20 | while read f; do echo "=== $f ===" cat "$f" done # Project instructions find ~/.claude/projects -name "CLAUDE.md" 2>/dev/null | head -20 | while read f; do echo "=== $f ===" cat "$f" done ``` **Cursor:** ```bash cat ~/.cursor/rules/*.md 2>/dev/null find . -maxdepth 3 -name ".cursorrules" 2>/dev/null | head -10 | while read f; do echo "=== $f ===" cat "$f" done ``` **Windsurf:** ```bash cat ~/.windsurf/rules/*.md 2>/dev/null find . -maxdepth 3 -name ".windsurfrules" 2>/dev/null | head -10 | while read f; do echo "=== $f ===" cat "$f" done ``` **Generic (AGENT.md / CLAUDE.md / rules files):** ```bash find ~ -maxdepth 4 \( -name "AGENT.md" -o -name "CLAUDE.md" -o -name ".cursorrules" -o -name ".windsurfrules" \) 2>/dev/null | head -20 | while read f; do echo "=== $f ===" cat "$f" done ``` After scanning, present what was found and proceed to Step 3. ### Step 3: Parse & Categorize Process the imported data (whether from ZIP, prompt response, or local scan) into these categories: **KEEP:** - Identity (name, profession, industry, language) - Writing style and communication preferences - Tools, platforms, tech stack - Active projects and workflows - Structural preferences (how to organize/present info) - "Don't do this" rules and corrections - Domain knowledge and expertise areas - Behavioral patterns and habits **FILTER OUT:** - Completed one-off tasks - Outdated context (finished projects, old deadlines) - Source-specific references ("as a ChatGPT user...") - API keys, tokens, passwords — **never migrate credentials** - Hallucinated or inaccurate memories (flag suspicious ones) - Duplicate or redundant entries ### Step 4: Review & Confirm Present the cleaned data to the user, organized by destination: **EN:** > 📋 **Migration Preview** > > **Writing to USER.md (your profile):** > - Name: ... > - Profession: ... > - Language: ... > - Communication style: ... > > **Writing to MEMORY.md (long-term memory):** > - [project/knowledge/experience entries...] > > **Writing to TOOLS.md (tool preferences):** > - Tools: ... > - Platforms: ... > > Anything to change? I'll write it once you confirm. **ZH:** > 📋 **迁移预览** > > **写入 USER.md（你的画像）：** > - 姓名：... > - 职业：... > - 语言偏好：... > - 沟通风格：... > > **写入 MEMORY.md（长期记忆）：** > - [项目/知识/经验条目...] > > **写入 TOOLS.md（工具偏好）：** > - 常用工具：... > - 平台配置：... > > 有什么要改的吗？确认后我就写入。 Wait for user confirmation. They can: - Approve all - Remove specific items - Edit entries - Add things that were missed ### Step 5: Write to Memory System After confirmation, write to the appropriate files: **USER.md** — Identity, communication preferences, language, timezone **MEMORY.md** — Knowledge, projects, experience, behavioral patterns **TOOLS.md** — Tools, platforms, environment-specific notes Rules for writing: - **Merge, don't overwrite** — if these files already have content, integrate new data with existing - **Preserve structure** — follow the existing format of each file - **Add a migration note** — append a comment like `<!-- Migrated from ChatGPT on 2026-03-30 -->` - **Use the user's language** — write entries in the language the user communicates in ### Step 6: Verify After writing, confirm: **EN:** > ✅ **Migration complete!** > > Here's what I now know about you: > [Brief summary of key imported info] > > This info is now in my memory system — I'll use it naturally in our conversations. > Feel free to tell me if anything needs updating. **ZH:** > ✅ **迁移完成！** > > 现在我知道的关于你的事： > [Brief summary of key imported info] > > 这些信息已经写入我的记忆系统，以后的对话中我会自然地使用它们。 > 随时可以告诉我补充或修正任何内容。 ## Special Cases ### Multiple Sources If the user used several AI assistants, handle them sequentially. Deduplicate across sources before writing. ### Partial Migration User might say "just bring over my writing preferences" — respect scope limits. Only migrate what they want. ### Conflict Resolution If imported data conflicts with existing memory (e.g., different profession noted), ask the user which is current. ### Re-migration If the user runs migration again later, merge new data with existing. Don't create duplicates. ## Platform-Specific Notes ### ChatGPT Data Export ZIP Structure ``` ├── conversations.json ← Main conversation history ├── model_comparisons.json ← Model comparison data ├── message_feedback.json ← Thumbs up/down data ├── shared_conversations.json ├── user.json ← Account info └── chat.html ← Rendered conversations ``` The parser (`scripts/parse-chatgpt-export.js`) extracts: - User messages patterns and topics - Correction patterns (user said "no, I meant...") - Frequently discussed subjects - Writing style from user messages - Tool/platform mentions (word-boundary-aware detection) - Project references ### Claude.ai Users can access their memory at claude.ai → Settings → Memory. They can either: 1. Use the prompt method (Step 2B) 2. Manually copy their memory entries ### Gemini Gemini stores "Saved Info" in Settings. Prompt method works best. ### Copilot Limited memory capabilities. Prompt method captures what's available. ## Security - **NEVER migrate API keys, tokens, or credentials** - Warn user if imported text contains what looks like secrets (regex: `/(?:sk-|ghp_|token|password|secret|key)\s*[:=]/i`) - All imported data is shown to user before writing — no silent imports - ZIP files are processed in temp directory with path traversal protection, cleaned up after