banana-claude-codex-import

Import local Claude Code (~/.claude/) and Codex (~/.codex/) conversation history into OpenClaw's memory system, making past coding work queryable and part of long-term memory.

When to use this skill

Use this skill when the user:

• Wants to import, migrate, or consolidate Claude Code or Codex sessions
• Asks to "导入" or "migrate" past coding agent conversations
• Wants to make past Claude Code / Codex work searchable in memory
• Needs to rebuild memory from previous coding agent sessions
• Mentions ~/.claude/ or ~/.codex/ session files
• Is reviewing past work and wants to extract key decisions/patterns into memory

What this skill does

1. Detect whether Claude Code and/or Codex data exists locally
2. Preview with --dry-run to show session count, date range, and sample content
3. Import all sessions to logs/message-archive-raw/
4. Generate memory candidates for distillation into MEMORY.md

Workflow

Step 1: Dry-run first

Always run with --dry-run first to verify data coverage:

python3 ~/.openclaw/skills/banana-claude-codex-import/scripts/import_conversations.py --dry-run

Check:

• Session count (aim for 50+ for Claude Code, 50+ for Codex)
• Date range
• Whether content looks real (not just noise/environment context)

If results look good, proceed.

Step 2: Full import

python3 ~/.openclaw/skills/banana-claude-codex-import/scripts/import_conversations.py

This writes:

• Session archives → logs/message-archive-raw/{claude_code,codex}/
• Memory draft → memory/YYYY-MM-DD_import_candidates.md

Step 3: Distill into MEMORY.md

Read memory/YYYY-MM-DD_import_candidates.md and extract high-value content into MEMORY.md:

## Claude Code / Codex 导入记忆（YYYY-MM）

> N sessions imported from `logs/message-archive-raw/`

### 技术决策
- [key decision 1]
- [key decision 2]

### 项目 / 主题
- [project 1]: [what was done]
- [project 2]: [what was done]

### 工具 & 工作流
- [tool/workflow pattern observed across sessions]

Source data

Source	Path	Format
Claude Code sessions	~/.claude/projects/*.jsonl	JSONL, user/assistant/tool_result entries
Claude Code index	~/.claude/history.jsonl	sessionId → metadata
Codex sessions	~/.codex/sessions/**/rollout-*.jsonl	JSONL, session_meta/response_item entries

CLI options

--dry-run              # Preview only, no files written
--source both          # Import both Claude Code and Codex (default)
--source claude_code   # Claude Code only
--source codex         # Codex only
--since YYYY-MM-DD     # Only sessions on or after date

What makes a good memory distillation

After import, the most valuable things to capture in MEMORY.md:

1. Technical decisions — architecture choices, tools selected/rejected, what failed
2. Recurring workflows — patterns in how the user works with coding agents
3. Project context — what each project is about, key milestones
4. Personal preferences — working style, communication patterns, explicit preferences stated

Avoid: verbatim copying of conversations. Capture the distilled meaning.