Browser-based graphical shell and control plane for a Linux server you own — real terminal, file manager, live metrics and BYOK AI in one mobile-first workspace. Next.js 16 · React 19 · node-pty · Tailwind v4 · shadcn/ui. Public alpha, self-hosted.
Memory preservation and second-brain visibility
Source protection
Age is not proof that a session was learned. The current preservation policy fails closed: unreviewed, pinned and reviewed session source files remain protected from automatic archive/artifact/action-index deletion. Review alone is not authorization to destroy source evidence. The record model supports review provenance, but there is no owner UI for releasing sources yet.
Compaction still archives sanitized raw context. Evicted action-index records are stored in bounded immutable segments before leaving the hot index. Lookup is bounded; persisted history is not a promise of an unlimited single response.
This deliberately grows storage until explicit release criteria and verified backup coverage are available. It is not a disk reclamation mechanism.
Cleanup
Settings > Cleanup requires an Owner session, a fresh five-minute server preview,
explicit category selection and confirmation. No category is preselected.
Broad /tmp age-only cleanup is blocked by the owner API. Trash and old logs
remain irreversible, and displayed estimates are not guaranteed freed bytes.
Primary session records, memory stores and volumes are not cleanup targets.
The underlying legacy Docker category still combines image and build-cache
pruning. It must not be described as build-cache-only. CLI/raw API mutations
must include the fresh preview_id, selected ids, and confirm: true.
Local server-memory backup
Settings > Backup separates Server memory backup from Browser backup. The server endpoint is owner-only. It accepts no arbitrary client source or destination paths and never returns backed-up contents.
mso memory-backup preview
mso memory-backup history
mso memory-backup history <offset> <revision>
mso memory-backup create --confirm
mso memory-backup verify <id> <manifest-sha256> --confirm
The allowlist covers configured typed memory, session archives, recipes,
organization data, canonical project .agent/knowledge sources, and sessions.
Core memory/project evidence precede the larger session inventory. Credential
stores, browser profiles, known config/lock/temp paths are excluded. Skill
catalogs, databases and every other VPS directory are not covered by this scope.
Snapshots are private local copies under ~/.mso/backups/memory/<id> with
compressed per-file checksums and a manifest checksum. Verification restores
into a NEW isolated directory and checks the bytes again; it never overwrites
originals. Missing sources, exclusions, rejected files and truncation are
reported. A partial snapshot remains partial after an integrity-successful
restore. Concurrent input files can change; this is a per-file verified copy,
not a transactional point-in-time snapshot, and it is not offsite protection.
Limits: 10,000 files, 40,000 scan entries, 1 GiB input, 32 MiB/file and a 45-second scan budget. Full resumable inventory/backup remains open. History pagination continues the saved-snapshot listing only; it does not resume backup copying.
Settings > Backup > Load saved server snapshots reads metadata in bounded pages of 12 snapshots, at most 200 directory entries and five seconds per request. Continuation uses a directory revision; changes require refreshing the first page. The listing is directory order, not a promise of chronological ordering. Corrupt metadata is shown as unreadable rather than hidden as empty. Manifest validation requires file and byte totals to match its entries before accepting coverage.
New successful restore rehearsals write an immutable receipt bound to the manifest checksum. History reports integrity at that recorded time, not a fresh integrity check. Older snapshots without receipts remain "not recorded"; they are not silently promoted or invalidated. A partial snapshot stays partial after a passing restore. Selecting history metadata never starts restore automatically. Receipt write failures surface as errors and preserve the isolated restore evidence. No backup result grants deletion permission.
Visibility and access
Workflows > Sources is owner-only discovery across graph principals. Metadata retains origin; existing client/MCP graph reads remain principal-private. A foreign workflow can only be copied explicitly into an inactive review draft with action nodes disabled and its source digest preserved. Source mutations and copying stale revisions are refused.
Session details > Self-improve fetches linked active/archived recipes, filters best/latest-run provenance before pagination, and distinguishes errors from empty results. Graph-cap/persistence warnings have durable receipts when the receipt store is writable. This is not proof that every historical session was learned or that every procedural-memory write succeeded.
Memory is pinned on the first home page. The owner projection includes bounded Organization, workflow, recipe and typed-agent metadata, source navigation and honest truncation notices. Private agent values are omitted. This does not widen retrieval grants for agents, operators or MCP clients.
Remaining integrations and release gates
JEV session preservation now writes a checksum-bound value capsule for every current session source. The capsule stores bounded semantic step summaries, unique tool routes, artifact references, decision metadata and the continuation packet; it does not store credentials or raw authorization material. agent_session_resume reuses the exact current capsule when its source SHA-256 and byte count still match, so even a session with zero probabilistic optimization recommendations retains a useful continuation asset. Capsule existence is preservation/value evidence, not raw-source deletion authority.
Jev now participates in automatic observed workflow-memory admission only. It may classify a bounded successful-route candidate as ignore, episodic, semantic, procedural, or candidate-recipe. Explicit user memory writes remain authoritative, candidate-recipe does not bypass existing recipe maturity/promotion rules, and a Jev decision never grants memory deletion or source-evidence cleanup authority.
New source behavior becomes active only after the normal verified build and release process. Passing unit/type/architecture tests alone is not production or browser acceptance. Keep tested, merged, deployed and live-verified distinct.