Data Migration

​

Overview

Myrm provides a complete data migration system that lets you transfer your AI assistant data from other platforms — including memories, skills, API keys, MCP configurations, and workspace rules — with full preview, conflict resolution, and one-click rollback. Supported Sources: The local and desktop builds auto-discover Hermes, OpenClaw, Claude Code, Codex, and Pi data on your machine; the cloud-hosted build recognizes an uploaded archive (ChatGPT exports, gbrain, and more). Memory Center can additionally import 18 ready sources (19 registered; MemWeaver is not ready yet), including Cursor, Windsurf, Trae, Mem0, Letta, LangChain, Zep, Hindsight, PLUR, AgentMemory, ChatGPT, and gbrain. Known boundary: Hermes session history can currently be imported only from exported JSON/JSONL files; the wizard does not read Hermes’s local state.db directly yet. Kanban boards are not migrated automatically.

​

Migration Wizard (3-Step GUI)

​

Step 1: Discover

Navigate to Settings → Memory Center → Migration to start the wizard.

• Local/Desktop: Click Scan to auto-detect installed AI assistants on your system
• Cloud: Drag and drop a .zip export file from your previous assistant

The scanner checks well-known directories (macOS, Windows, Linux) and reports what it found with confidence scoring (high/medium/low).

​

Step 2: Preview (Dry-Run)

Click Preview on any detected source to see exactly what will be imported — without writing anything:

• Coverage Matrix: Shows which data categories are ready, need review, or require manual action
• Migration Lanes: Persona → Agent, Facts → Memory, Skills → Review Queue, Keys → Explicit Confirm, MCP → Audit
• Token Economics: Compare token usage before and after migration
• MCP Server Preview: See which MCP tools will be available after import
• Conflict Detection: Highlights items that would overwrite existing data

​

Step 3: Confirm & Import

Choose a target Agent (or create a new one), then confirm the import:

• Secrets are only imported with explicit opt-in (never silently)
• Skills go to a review queue — you approve each before it activates
• A transaction batch ID is generated for every import
• ZIP skill migration uses a stable two-phase error contract (message + error_code) in both preview and confirm, so UI remediation stays deterministic.

​

Post-Import Readiness (First Chat Guidance)

After a successful import, Myrm live-resolves readiness before your first message — you are never left guessing why chat feels broken. Design principles:

• Soft gate — the assistant can still reply; you get guidance, not a hard block
• Issue-aware CTA — each gap links to the exact settings page (settings_path SSOT)
• Wizard recheck — Migration Result step can re-run readiness before “Start chat”
• Queue anchor — handoff state survives navigation until first chat consumes it

Verified (2026-07-29): Chrome E2E sequential 3-leg signoff 3/3 PASS (mcp_warning · provider_critical · diagnostic_critical) under parallel wave (≤8 leases, no stopping other pytest) · 65+ unit/integration tests · competitors typically silent-fail or CLI-only with no post-import UX. Additional readiness checks after Hermes migration may surface voice or MoA advisor overlay setup hints when those capabilities were used in Hermes but are not configured in your Myrm Agent profile — each links to the relevant Settings page without blocking chat.

​

Hermes Cron Job Import (Paused by Default)

When migrating from Hermes, the wizard can import scheduled jobs from cron/jobs.json: Honest boundaries:

• Hermes model fields are not migrated — jobs use the target Agent’s default model
• Kanban boards are not auto-migrated (coverage matrix shows kanban_not_migrated)
• No Chrome MCP E2E sign-off for cron import yet — backend covered by test_hermes_cron_converter.py 7/7 passed (Aug 2026)

No competitor (Hermes CLI claw migrate, OpenClaw, Pi, deer-flow, LobsterAI, CoPaw, jiuwenclaw) offers a GUI migration wizard with cron preview + paused import + rollback.

​

Pi Migration (5-Lane Mapping)

When migrating from Pi, the wizard auto-scans ~/.pi/agent/ and provides five-lane mapping: Safety guards:

• Unknown session versions (≠ v3) are auto-skipped silently
• Empty messages, invalid JSON, non-dict content blocks are safely skipped
• Pi has no MCP configuration (tech stack difference) — does not affect migration completeness

Verified (2026-08-03): test_pi_migration.py 25/25 PASS (covering discovery / loading / session parsing / skills / credentials / edge cases — 6 categories, 25 scenarios)

​

Hermes MoA Reference Import (Wizard Confirm)

When your Hermes config.yaml includes a moa block with reference models, the migration wizard can write them into the target Agent’s engine_params.moa_overlay on confirm: Honest boundaries:

• Only the default Hermes preset’s reference set is imported (multiple presets with different ref pools → create multiple Agent Profiles)
• Hermes aggregator model is not migrated — Myrm primary model absorbs advisor output (by design)
• Provider keys must be configured in Myrm before refs resolve; unresolvable refs are skipped with logging

Verified (2026-08-03): test_hermes_moa_migrator.py 19/19 + confirm integration path in import archive — part of 30/30 MoA server pytest suite.

​

Zero-Friction Cross-Platform Session Migration (Claude Code / Codex / Cursor)

When migrating from Claude Code, OpenAI Codex CLI, or Cursor to Myrm, the wizard not only transfers rules and configs, but also provides high-fidelity conversational transcript & context handoff: Honest Boundaries:

• Transcripts are ingested into structured canonical turns bound to the target Agent Profile and persisted in SQLite;
• Single ultra-long lines trigger character-level safe truncation to guarantee context protection;
• Side effects of past tool calls are not re-executed against real external environments.

Machine Verification Evidence (2026-09-28):

• Harness Framework Unit Tests: 99% coverage (tests/runtime/context/test_transcripts_parser.py 8/8 passed);
• Server Migration Suite: Full suite 232/232 PASS (tests/services/migration/);
• Real Browser Chrome MCP E2E: tests/e2e/test_competitor_session_migration_chrome_e2e.py 1/1 PASS (17.14s, zero mock, verified via real WebUI DOM upload and review modal).

​

Rollback

Every import can be fully reversed:

1. Go to Settings → Memory Center → Recent Imports
2. Click Rollback on any import batch
3. Preview what will be undone (rollback dry-run)
4. Confirm to restore your previous state

​

Backup & Remote Sync

​

Whole-App Backup and Restore

Open Settings → My Workbench → Backup and restore to move everything in the app — chats, agents, skills, memory and settings — to another device, or to keep a safe copy:

• Create a backup: packs your data into one file. Model keys and connections stay out of it unless you include them, sealed with a passphrase only you know; if you forget the passphrase, those keys can’t be recovered
• Check before anything changes: restoring first verifies the file and shows what it contains and what it will replace. If some items couldn’t be read when the backup was made, you are warned that they are not included
• Finish on restart: the restore is applied while the app restarts, before your data is opened
• Undo: Undo restore puts everything back the way it was before the last restore
• Automatic backup: saves a full copy on a schedule to a folder you choose. Automatic backups are ordinary, unencrypted files and never include logins, model keys or connections
• New device: install Myrm, open the same page and pick your backup file — nothing needs to be set up first. Hosted workspaces also offer Restore from a backup on the first screen of their welcome wizard
• Before updates: a database snapshot is taken before each update, so a failed update can be rolled back

​

Local Backup

Create manual backups of all memories at any time:

• Settings → System → Backup → Create Backup
• Stores: all memory types, shared contexts, conversation history
• List, restore, or delete backups as needed

​

Remote Backup (S3 / WebDAV)

Back up your memories to storage you own, on a schedule or whenever you like (Settings → Memory Management → Cloud Sync):

• S3-compatible: Any provider (AWS S3, Cloudflare R2, MinIO, etc.)
• WebDAV: NAS devices, Nextcloud, ownCloud, etc.
• What is covered: the memories of your default agent and the ones shared by all agents. Memories kept only by other agents, projects or chats are not included, and neither are conversations, skills or settings — use the whole-app backup above for those
• Not encrypted: backups are ordinary files, so use a location only you can access
• Schedule: hourly, every 6 hours, every 12 hours or daily. It runs only while Myrm is open and starts within about an hour of the chosen time; if backups keep failing, you are notified
• Retention: the newest backups made by this app are kept (10 by default) and older ones are deleted. Backups from your other devices, and any other files in the folder, are never touched, so several devices can share one folder
• Restore: pick a backup from the list; it is merged into your current memories and nothing is deleted

​

First Launch

Local and desktop builds open straight on the home page, with no setup wizard in the way. If Myrm finds Hermes, OpenClaw, Claude Code, Codex, or Pi data on your machine, a Discovery Banner appears under the Notices button on a new chat, with a shortcut into the migration wizard (Settings → Memory Center → Migration) — so your first conversation already has your personality, memories, and tools. Hosted workspaces start with a first-run wizard whose first step is the same one-click migration.

​

Export

​

Full Archive Backup & Restore

Export your entire Myrm workspace as a structured archive that preserves all data relationships:

• Conversations: Full message history including compaction state (compressed summary), session notes, tool usage stats, and token cost summaries
• Memory: All memory types (episodic, semantic, procedural, preferences, shared context)
• Agent Configurations: Complete Agent profiles (model, prompt, skills, MCP bindings) via JSON round-trip export/import
• Section-Selective Restore: Import the full archive or choose specific sections (conversations only, memory only, etc.)

This is the recommended path for disaster recovery, cross-instance cloning, or team onboarding.

​

Cross-Deployment Migration (Local ↔ Cloud)

Migrating between Local WebUI, Tauri Desktop, and Cloud-hosted goes through a backup:

• Backup and restore: Create a backup on one deployment and restore it on another (see above). Settings that describe the machine itself — security rules, approvals, paired channels — stay as they are on the target, and saved secrets that can’t be decrypted elsewhere are entered again
• Archive Route: Export a full archive from one deployment, restore on another with section selection

Don’t copy individual database files between deployments: your data lives in several databases and files, and part of it is bound to the machine it was created on. Among the open-source agents we inspected, we did not find an equivalent end-to-end backup/restore path across local, desktop, and cloud; most are designed around a single deployment mode. That is an observation from the sources we read, not a claim that no competitor could add it later.

​

Memory Archive Export

Export your Myrm memories as a structured JSON archive (myrm_memory_archive format) that can be imported into another Myrm instance or shared.

​

Dataset Export

Export conversation data in training-ready formats:

• ShareGPT: Standard fine-tuning format
• PII Redaction: Optionally strip personal information
• Quality Filtering: Min turns, min content length, success-only
• Incremental: Export only new data since last export

​

Data Root Path Migration Wizard & Storage Governance

As projects accumulate over long operational cycles, session memory, vector indexes, snapshots, and artifacts inevitably expand. Myrm provides an enterprise-grade data root path migration wizard and multi-dimensional storage governance system:

​

Core Security Mechanisms & Safeguards

1. Triple Hard Safeguards:
   • Anti-Self-Nesting Check: Strictly prohibits migrating the root data path into the current data directory or any of its subdirectories, eliminating infinite recursion cycles during directory moves.
   • 1.2x Available Space Headroom Preflight: Accurately measures destination partition free capacity before migration; requires at least 120% of source size, preventing disk-full SQLite WAL tearing during cross-volume moves.
   • Destination Conflict Isolation: Safely blocks and isolates migration if the target path already contains valid data, preventing accidental overwrites.
2. Two-Phase Sensitive Action Ticket:
   • Requires short-lived, single-use cryptographically signed tickets. Users must run Dry-Run preflight and explicitly confirm before execution, preventing CSRF hijacking and UI accidental triggers.
3. Running Background Task Awareness & Exclusive Write Lock:
   • Dynamically monitors active background long-running tasks (e.g., persistent shell scripts, evaluation runs). If tasks are running, immediately rejects the migration with dynamic active task count feedback (HTTP 409 Conflict), preventing mid-task write tearing;
   • Acquires an exclusive write lock during migration and triggers WAL TRUNCATE flushing to ensure seamless database cutover.
4. Sub-Second Atomic Rollback:
   • Employs granular transactional compensation. If interrupted by cross-filesystem (EXDEV) copying errors or unexpected power failure, the system rolls back source directory state and removes partial destination artifacts in sub-seconds (zero corruption, zero leftovers).
5. Storage Profiling & Online Compaction:
   • Breaks storage down by category (chat and settings database, conversation state, agent workspaces, search index, long-term memory, activity logs, database backups, snapshots, and everything else); a background job drops superseded conversation checkpoints and shrinks the checkpoint database once enough space is free, and the chat database can be compacted on demand while the app keeps running.

​

Comparison with Competitors

How to read this table: ✅ means we found and verified the capability in the product or source we inspected; ❌ means we did not find it in the scope we checked, not proof that it is impossible. The comparison is a snapshot, and competitors change. Comparison targets: Hermes, OpenClaw, Codex, DeepSeek-Harness, Pi, OpenCode.

​

Coming from WorkBuddy or Tencent WeKnora?

There is no one-click importer yet (WorkBuddy is closed-source with no export format; WeKnora stores data in its own Go/Postgres stack) — but every workflow has a counterpart in Myrm. Where one is still partial the row says so, and the gaps are listed under the table: Honest gaps (in progress, not hidden): one-click prompt enhance, output style presets (table-first, no preamble), Ask-only mode, agent-authored wiki generation with version rollback, and the extended connector matrix (Feishu Drive structure-preserving sync, Notion, GitLab, DingTalk docs, RSS). What will not come: multi-tenant Workspace RBAC (conflicts with our single-user-per-sandbox architecture) and a CLI product (we are GUI-first).