The reverse of export: read an archive, verify it, and recreate the workspace. It's what makes the export a genuine round trip rather than a one-way dump.
What it restores
Ids and timestamps are preserved, which is the part that matters. Citations keep resolving, [[wikilinks]] keep pointing at the right memories, and "published 3 July" still says 3 July. An import that renumbered everything would technically hold your data and practically destroy its usefulness.
Restored: memories with their projects and types, collections with every artifact version, comments and approvals pinned to their versions, tag definitions with their privacy settings, and notification rules.
Not restored, matching what the export deliberately leaves out: secrets, API tokens, share codes, entitlements and inbox items. Regenerate the credentials; they were never in the archive.
People are references, not accounts. Author ids map through the archive's users.yaml, defaulting to the admin doing the import. Attribution labels are always preserved, so the history still reads correctly - "written by Ana's instance" stays true on the page even where the underlying account doesn't exist in the new install.
It goes into an empty workspace
Import targets a fresh, empty workspace. If any id in the archive collides with something already present, the import refuses rather than silently remapping ids and quietly corrupting the links.
Merging an archive into a populated workspace isn't supported yet. Restoring into a new workspace and moving what you want across is the supported path.
Doing it
Settings -> Backups -> import, upload the archive.
From the CLI, with a dry run first - worth doing every time:
kythe import --archive ./workspace.tar.gz --dry-run
kythe import --archive ./workspace.tar.gz --into <target-space>The archive is verified before anything is applied: the format version is checked (a newer format than the install understands is refused), and every file's sha256 is checked against the manifest. A corrupt or truncated archive fails up front rather than half-importing.
Afterwards
Memory full-text search works immediately. Semantic vectors for memory backfill on the next boot pass, and the search index over published content is rebuilt by a background reindex - so recall gets steadily better over the first few minutes rather than being wrong.
Moving between hosted and self-host
Export from one, import into the other. The format is the same in both directions, so it works hosted -> self-host, self-host -> hosted, and self-host -> self-host. Useful for a trial that turns into a deployment, or a self-host install that needs to move box.