# Import & export

> Bring the memory you already have — and take all of it with you, any time.

Cold start is optional. Import what you already have and the Dreamer
consolidates it on its first night — a wiki instead of an empty screen.
And everything that lives here leaves whole: one click exports your
memory as files you own.

## Import from files

Open [app.oceandb.ai/dashboard/import](https://app.oceandb.ai/dashboard/import).

**Markdown / Obsidian.** Choose files or a whole vault folder. Each file
becomes an entry — the title comes from frontmatter or the filename, a
frontmatter block is stripped from the content. Re-importing skips what
is already there, so running it twice is safe.

**ChatGPT.** Request your export in ChatGPT (Settings → Data controls →
Export data), then upload the zip it emails you — or just its
`conversations.json`. Each conversation becomes one entry with the full
back-and-forth. Parsing happens in your browser; large archives take a
minute.

Imported entries are embedded overnight by the maintenance sweep, so
semantic recall finds them from the next day; text search sees them
immediately.

## Import through your agent

Memory that lives in `CLAUDE.md`, Claude Code auto-memory, or scattered
notes doesn't need an uploader — your agent reads it and remembers.
With the OceanDB MCP [connected](/docs/connect), paste this into
Claude Code from your project directory:

```
Read CLAUDE.md and any memory/notes files in this project
(~/.claude/CLAUDE.md too). For each durable fact, decision, or
preference worth keeping, call the OceanDB remember tool — one
memory per call, a clear title, the fact as content. Skip
boilerplate and anything transient. Report what you remembered.
```

`remember` embeds each memory as it lands, so these are searchable
immediately.

## Export

From [the same page](https://app.oceandb.ai/dashboard/import), two
downloads:

- **Markdown zip** — every entry and wiki page (archived included) as
  `.md` files with frontmatter. Readable anywhere, greppable forever.
- **SQL dump** — entries, pages, links, insights, projects, and tags as
  plain `insert` statements, ids preserved so the citation graph stays
  coherent.

No lock-in is a feature of the memory, not a promise on a pricing page.