SKILL.md
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
- Detect whether Claude Code and/or Codex data exists locally
- Preview with
--dry-runto show session count, date range, and sample content - Import all sessions to
logs/message-archive-raw/ - 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-runCheck:
- 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.pyThis writes:
- Session archives →
logs/message-archive-raw/{claude_code,codex}/ - Memory draft →
memory/YYYY-MM-DDimportcandidates.md
Step 3: Distill into MEMORY.md
Read memory/YYYY-MM-DDimportcandidates.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, sessionmeta/responseitem 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 dateWhat makes a good memory distillation
After import, the most valuable things to capture in MEMORY.md:
- Technical decisions — architecture choices, tools selected/rejected, what failed
- Recurring workflows — patterns in how the user works with coding agents
- Project context — what each project is about, key milestones
- Personal preferences — working style, communication patterns, explicit preferences stated
Avoid: verbatim copying of conversations. Capture the distilled meaning.