Devices and portable project paths

After workspace sync or export/import, memories, decisions and conversation chunks remain usable across devices. Device metadata describes origin; it does not filter retrieval or create another knowledge base. Both sides still need the same embedding profile for vector sync. Recording device origin does not automatically transfer data; use the existing sync workflow.

New memories record a stable device ID, hostname, OS and recorded time. Files also record source device and project-relative path. Graph facts retain origin. Stream/snapshot sync and workspace copies preserve it. Old records show Unknown origin (legacy) instead of being attributed to the receiver. Reindex local files to refresh provenance; metadata itself does not require re-embedding.

imprint device
# On device B, map the canonical stored project name to its local checkout:
imprint workspace roots add /code/project --workspace work
imprint workspace map project /code/project --workspace work

Normal file ingestion sets the project mapping automatically. A mapping changes only this device’s workspace policy; root approval/sharing rules still apply. Maps are not synced. A file’s project-relative path resolves under the mapping of the project it was indexed in (its origin_project), so a file moved to another project (ADR 0036) still resolves to where it lives.

A memory can reference C:\Users\alice\project\src\app.py from A while B uses /code/project. Imprint preserves the original as a reference and resolves src/app.py through B’s approved map. Retrieval adds path_context: current device/cwd, original path, status and optional local_source_path. A hint requires an existing file inside both the approved root and current process cwd. Traversal/symlink escape produce no hint. Hashes describe indexed bytes; verify current files before edits.

Without a mapping, or outside cwd, shared content remains available. Original paths are not instructions to access local files. Conversation text and example commands are preserved, including historical paths. MCP advertises device/cwd on initialization and exposes device_context; its cwd comes from the coding host. Restart the connection from the project directory if needed. Dashboard chat also receives current device/cwd and guidance to treat historical paths as references; its cwd belongs to the Imprint server. /api/device and dashboard memory/source detail expose the same context.

Imported paths are unverified locally even when their strings match local paths. Imported URLs are not re-fetched here: URL refresh only follows URLs this device recorded. Refresh/pruning do not delete foreign or unknown sources based on local filesystem state. Physical-path purge uses locally recorded provenance; use source-key purge for synced/legacy records. Snapshot restore refuses to bypass purge exclusions: use a fresh workspace or explicitly restore exclusions first.

data/device_id.txt is shared by Go/Python sync; existing IDs stay stable and concurrent startup provisions one ID. Hostname is a captured human-readable label. The ID is not cryptographic authentication. Bundles do not copy device identity or maps. Whole-data-directory clones require a distinct local identity before recording more data; prefer supported export/import or peer sync.

See workspaces, MCP and ADR 0017.