#What lands in your downloads folder

Three different layouts, depending on what you exported and to which format.

• A single conversation
• A whole library
• Filenames

#A single conversation

#No attachments → a bare file

How does the RPC auth work - 2026-07-31T06-40-38.md

There is nothing to bundle, and wrapping a lone .md in a ZIP is friction for no benefit.

#With attachments → a ZIP

How does the RPC auth work - 2026-07-31T06-40-38.zip
├── conversation.md
└── attachments/
    ├── network-log.har
    └── screenshot of failing request.png

The document is always named conversation.<ext> inside the archive; the archive itself carries the title. Every link in the document is relative (attachments/…), so extract the ZIP before following links. Most PDF readers and browsers will not resolve a relative link out of a compressed archive.

CSV and JSONL never produce a ZIP. Neither format has a way to reference a file on disk, so shipping the bytes beside them would be dead weight.

#Colliding filenames

Two attachments called screenshot.png become screenshot.png and screenshot (2).png. The suffix goes before the extension so the file still opens in the right application.

#A whole library

#Per-conversation formats → one folder each

Markdown, plain text, JSON, HTML, PDF, Word and Obsidian:

ai-studio-library-2026-07-31T06-40-38.zip
├── index.md
├── manifest.json
├── how-does-the-rpc-auth-work/
│   ├── conversation.md
│   └── attachments/
│       └── screenshot.png
└── another-conversation/
    └── conversation.md

index.md is a readable listing with a link to each conversation and a section for anything that failed. manifest.json is the same information for tooling; fields are documented in data reference.

Folder names come from the conversation title. Two conversations with the same title get the second one's id appended: my-notes (1AbCdE).

#Combined formats → one file

CSV and JSONL produce a single file, not an archive:

ai-studio-library-2026-07-31T06-40-38.csv
ai-studio-library-2026-07-31T06-40-38.jsonl

The CSV's header row appears once at the top, not once per conversation.

#Large libraries

A per-conversation export splits into multiple archives once the next conversation would push one past 256 MB (measured before compression, since that is what predicts memory pressure), so the finished files are smaller.

ai-studio-library-2026-07-31T06-40-38 (part 1 of 3).zip
ai-studio-library-2026-07-31T06-40-38 (part 2 of 3).zip
ai-studio-library-2026-07-31T06-40-38 (part 3 of 3).zip

A part only exceeds 256 MB when a single conversation does on its own, and that conversation gets a part to itself rather than dragging its neighbours over the limit with it.

Every part lists the whole library. Its index.md names every conversation in the export: the ones in that archive are links, and the rest say which part holds them.

- [Weekly Notes](weekly-notes/conversation.md): 2026-07-14 · 42 messages
- Design Review: 2026-07-11 · 210 messages · **in part 3**

So whichever part you open, nothing in the export is unaccounted for and no link points at a file that is not there. The failure list appears in every part, for the same reason. manifest.json mirrors this: conversations is what this archive holds, library is everything, with a volume number on each entry.

Extract the parts side by side and they merge into one folder.

Chrome may ask once whether to allow multiple downloads.

#Partial failure

One unreadable conversation does not lose the rest. Failures are collected and reported at the end, and recorded in manifest.json with the id, title and error message. The export only aborts if every conversation failed.

#Filenames

Conversation titles are sanitised for the filesystem: characters illegal on Windows are removed, the result is truncated to 120 characters, and names that collide with Windows reserved device names (CON, PRN, NUL, AUX, COM1…) get an -export suffix. An empty result falls back to conversation.

Inside an archive, names are shorter still, so that no path gets near Windows' 260-character limit. Windows Explorer refuses to open a ZIP at all if any one path inside it reaches that limit, and extracting has to fit the destination folder within it too.

A name cut short loses the end of its text, never its extension, and never ends in a space, dot or dash, which Windows would strip. The JSON export's name field still records the attachment's full original name.

The timestamp is the export time in UTC, ISO 8601 with : and . replaced by -, which sorts correctly in a file listing.