# Export AI Conversations

> Export locally stored AI assistant conversations and prompts from Antigravity, Codex, DeepSeek, WorkBuddy, and Cursor into Markdown with a unified index. Use for manual or scheduled daily backups; do not use for cloud-only histories that are not present on disk.

- Skill: `marvin-lijunzan/export-ai-conversations` (Agent Skill, multi-file: 7 files)
- Install (CLI): `npx skillmds@latest add marvin-lijunzan/export-ai-conversations`
- Raw SKILL.md: https://api.skillmd.com/api/skills/marvin-lijunzan/export-ai-conversations/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: DevOps & Infra
- Author: Marvin-LiJunZan (https://skillmd.com/u/marvin-lijunzan)
- Updated: 2026-09-21
- Page: https://skillmd.com/skills/marvin-lijunzan/export-ai-conversations

---


# Export AI conversations

Use the bundled script for the actual export. The script is deterministic and cheap to run: it reads local files and databases, converts recognized records to Markdown, and updates one index. Do not ask a language model to reformat every conversation.

## Default run

Run from the repository root:

    python skills/export-ai-conversations/scripts/export_ai_conversations.py

The default output is a chats directory under the current working directory. Override it when needed:

    python skills/export-ai-conversations/scripts/export_ai_conversations.py --output ".\conversation-exports"

The script discovers local application data from the current Windows user profile. Set an environment variable only when an application uses a non-default location:

    $env:ANTIGRAVITY_BRAIN_DIR = "$env:USERPROFILE\.gemini\antigravity\brain"
    $env:CODEX_HOME = "$env:USERPROFILE\.codex"
    $env:DEEPSEEK_DATA_DIR = ".\chats"
    $env:WORKBUDDY_DB = "$env:APPDATA\WorkBuddy\codebuddy-sessions.vscdb"
    $env:CURSOR_WORKSPACE_STORAGE = "$env:APPDATA\Cursor\User\workspaceStorage"

## Scheduled use

Schedule the command with Windows Task Scheduler, preferably once per day after the assistants have finished writing their local stores. Use python as the program and pass the repository-relative script path as the argument. Set the task to run whether or not the user is logged on only if the local application data is available to that account.

For a scheduled run, report only the counts and warnings printed by the script. A zero count for a source means that source was not found or had no parseable records; do not claim that its cloud history was exported.

## Operational rules

- Preserve existing files in the output directory. The script overwrites only files with its own stable prefixes and the generated index.
- Prefer transcript_full.jsonl over transcript.jsonl for Antigravity.
- Treat local stores as potentially changing while an app is open. If a database is locked or a JSONL line is incomplete, skip that record and report a warning; do not damage or rewrite source stores.
- Keep tool output bounded in Markdown so a single huge command result cannot make the daily archive unusable.
- This is a local-backup workflow. Do not upload conversation contents, call cloud APIs, or delete source data.

## Discovered local sources

- Antigravity: USERPROFILE/.gemini/antigravity/brain/*/.system_generated/logs/transcript_full.jsonl or transcript.jsonl
- Codex: CODEX_HOME/sessions, CODEX_HOME/archived_sessions, and CODEX_HOME/session_index.jsonl
- DeepSeek: DEEPSEEK_DATA_DIR/conversations.json or deepseek_data-*/conversations.json; otherwise the script checks ./chats, the repository chats directory, and USERPROFILE/chats
- WorkBuddy: APPDATA/WorkBuddy/codebuddy-sessions.vscdb
- Cursor: APPDATA/Cursor/User/workspaceStorage/*/state.vscdb

The source locations are intentionally user-relative or configurable. The script does not assume a particular Windows username or clone directory.