Skip to main content
Stateful work includes several data classes. Before choosing retention behavior, identify whether you need files, conversation, execution history, or process memory.

Lifecycle effects

Retention assumes the relevant storage remains available and has not been separately deleted. Client disconnect alone neither deletes these data nor cancels a Run. Database loss and key loss are distinct failures: the database backup contains history and artifacts, while the matching key enables protected credential recovery. Either one alone is insufficient to restore the complete workspace service. A project file is not automatically an Artifact. Export deliverables into the Run artifact directory and verify metadata, size, and hash. Later Runs do not overwrite earlier exported snapshots. File retention alone does not prove conversation recovery. Test a known file and a prior conversation fact separately. Temporary directories are not portable storage guarantees; Agent37 clears /tmp on its recovery path.

A concrete two-Run project

The first Run modifies app.js in the workspace and copies its deliverable version into that Run’s artifact directory. Only the copy exported by Runtime, validated and captured by the control plane becomes an independently downloadable Artifact. The second Run modifies workspace app.js again without changing the first Run’s Artifact; it must export its own version. Store each Run ID, Artifact ID and digest rather than identifying a version by filename alone. Before closing the Session, confirm that every file you need appears in the artifact list, then download it and verify its digest. Seeing a working preview or a file in the workspace does not prove that source code will remain retrievable after closure. Downloading the captured Artifact again after closure and confirmed cloud cleanup verifies that delivery no longer depends on cloud resource lifetime.

Run output, exports and previews are separate results

Run text output is not a complete workspace backup. On non-interrupted Runtime terminal paths, the current worker attempts artifact export before finalizing Run state. A failed or cancelled Run can therefore still have captured files. Conversely, the Runtime interrupted path skips that export attempt; uncaptured files cannot be assumed downloadable. Rejected export validation records artifact_export_rejected, so Harness exit alone does not prove successful delivery. A preview URL grants access to a running service; it is not a static archive URL. Expired authorization, Session closure or a stopped service process can each make the old URL unusable. For lasting presentation, deliver complete deployable artifacts to your own hosting system; keeping a temporary Hexagon preview is a different lifecycle.

Verify the retention you actually need

For the selected combination, check four separate outcomes: a later Run reads a modified file; it recalls a prior conversation fact; the preview service is available or deliberately restarted after planned wake; and exported artifacts remain downloadable with matching digests after closure. None proves the others. Separately verify historical events, artifact bytes and credential decryption in an isolated database restore, following backup and restore. The current local offline backup requires all Sessions to be closed, so it is not an active Sandbox snapshot tool. These are verification requirements, not new acceptance results claimed by this page. Export ordering sources: worker, Runtime export. For continuing edits, retain a Session and verify the idle strategy. For delivery, download artifacts or store an authorized application copy. For archival, manage database capacity, backups, and retention policy; current retention is not an unlimited-duration SLA. For a running service, manage process and preview authorization lifecycle rather than saving a URL alone. Sources: artifacts, serverless strategies. Next: downloads, multi-turn verification, backup/restore.

Evidence for retention, with its limits

The 2026-11-20 documentation journeys used Vercel Sandbox, built-in Pi 1.5.2 and the configured Anthropic-compatible Messages endpoint with deepseek-ai/DeepSeek-V4-Flash. Python and TypeScript ran independently. These records demonstrate specific retention outcomes; they are not a retention-duration guarantee or a result for every combination. The project evidence separates two facts that a single “success” label would hide: exported files remain downloadable while the preview is no longer authorized after closure. Its second-turn result also does not prove a universal memory service or that every Harness conversation format survives every Provider restoration strategy.

A missing item has different recovery sources

If the cloud workspace disappears but the control-plane database remains, first list the already captured Artifacts and retained Run output. Reconstruct a new project from those deliverables and any separately retained source inputs; do not present that as restoring the original process, native conversation or unsaved workspace changes. The initial upload is a baseline, not a continually synchronized copy of the project. If PostgreSQL is lost while a cloud process still runs, the process is not a replacement for Session ownership, event history, resource-operation receipts or stored Artifact bytes. Protect the remaining resource information and follow a deliberate database recovery procedure before issuing new work. Do not blindly create replacement resources while the prior resource outcome is unknown. If the database survives but its encryption key is lost, plaintext history or stored Artifact bytes do not reconstruct the encrypted connection credentials and invocation material. Conversely, a key alone cannot reconstruct the database. The local backup implementation bundles the schema dump with matching private identity material and verifies identity/decryption; it requires stopped local services and closed Sessions. A verified backup bundle is still different from an executed restore rehearsal. See backup and restore before choosing an operational recovery plan.