Continuity is not metadata — it's substrate. Without persistent continuity, an agent is not continuous.

A structured action logging system with integrity verification for OpenClaw agents. Uses lifecycle interception hooks (pr-12082) to provide durable audit trails, cryptographic hash chains, and recovery mechanisms for critical actions.

• 🔒 Pre-execution Logging - Critical actions are logged BEFORE execution using before_tool_call hooks
• ✅ Integrity Verification - SHA-256 hash chaining for tamper detection
• 💾 Pre-compaction Checkpoints - Capture state before memory context loss
• 🔄 Session Continuity - Automatic restoration of context on restart
• 🔍 Implicit Session Resumption - Detects and restores context when gateway restarts (even without explicit resume)
• 📊 Health Monitoring - Built-in health checks and diagnostics
• ⚙️ Configurable Log Levels - off | judgment | everything
• 🛡️ Fail-closed Security - Block actions if persistence fails (optional)
• 📝 Bidirectional Logging - Log both human inputs and agent responses

# Copy the plugin to your OpenClaw plugins directory
cp -r continuity-plugin /path/to/openclaw/src/plugins/continuity

Add to your OpenClaw configuration (e.g., openclaw.config.js or via CLI):

plugins: [
  {
    id: "continuity",
    path: "./src/plugins/continuity",
    config: {
      logLevel: "everything",
      storagePath: "~/.openclaw/continuity",
      enableIntegrityCheck: true,
      enablePreCompactionCheckpoint: true,
      blockOnPersistenceFailure: true,
    }
  }
]

Or use the CLI:

openclaw plugin install /path/to/continuity-plugin

⚠️ CRITICAL: Continuity data contains private information and must NEVER be committed to git.

# Add to .gitignore
echo ".openclaw/continuity/" >> .gitignore
echo "action-stream*.jsonl" >> .gitignore

• off - Disable all logging
• judgment - Log only decisions, analysis, and critical actions
• everything - Log all actions and messages (default)

When the OpenClaw gateway restarts, it typically creates a fresh session with resumedFrom: null. Normally this means all context is lost. However, the Continuity Plugin can detect implicit resumption by checking if there's been recent activity within a configurable time window.

How it works:

1. On agent start, the plugin checks the timestamp of the last recorded action
2. If the gap is less than implicitResumeThresholdMinutes (default: 30 min), it treats this as a continuation
3. Recent activity context is automatically restored and made available to the agent
4. A continuity_implicit_restore action is logged for audit purposes

Configuration:

{
  implicitResumeThresholdMinutes: 30  // Adjust based on your use case
}

Set to 0 to disable implicit resumption entirely.

The plugin registers handlers for these pr-12082 lifecycle hooks:

The key principle: actions with side effects are not "done" until persisted.

User Request
    ↓
Agent decides to use tool
    ↓
before_tool_call hook fires
    ↓
Action is logged to disk (SYNC)
    ↓
Log success? ──No──→ Block action (if fail-closed)
    ↓ Yes
Tool executes
    ↓
after_tool_call hook fires
    ↓
Result logged

~/.openclaw/continuity/
├── action-stream-YYYY-MM-DD.jsonl    # Daily append-only logs
├── .state.json                        # Stream state (sequence, last hash)
├── COMPACTION_MANIFEST.json           # Recovery manifest
├── EMERGENCY_RECOVERY.jsonl           # Write failure fallback
├── checkpoints/                       # Pre-compaction checkpoints
│   └── checkpoint-{id}.json
└── backups/                           # Automatic backups
    └── action-stream-{timestamp}.jsonl

Each action is stored as JSONL with integrity:

{
  "id": "action-uuid",
  "sequence": 47,
  "timestamp": "2026-02-09T18:48:00.000Z",
  "type": "tool_call",
  "severity": "critical",
  "platform": "openclaw",
  "description": "Tool call: write",
  "toolName": "write",
  "toolParams": { "path": "/file.txt", "content": "..." },
  "sessionId": "session-abc123",
  "_integrity": {
    "hash": "sha256_hash_of_content",
    "previous": "hash_of_entry_46"
  }
}

On startup, the plugin performs:

1. Disk space check
2. Stream integrity validation
3. Write access verification
4. Checkpoint availability check

╔════════════════════════════════════════════════╗
║         Continuity Status Report               ║
╠════════════════════════════════════════════════╣
║  Overall Health: ✓ HEALTHY                     ║
╠════════════════════════════════════════════════╣
║  Statistics:                                   ║
║    Total Actions: 1523                         ║
║    Storage Used: 2.45 MB                       ║
║    Disk Free: 45678 MB                         ║
║    Integrity: ✓ Valid                          ║
║    Checkpoints: 12                             ║
║    Last Action: 5m ago                         ║
╚════════════════════════════════════════════════╝

If memory is compacted and context is lost:

1. Check COMPACTION_MANIFEST.json for last checkpoint
2. Restore from checkpoint in checkpoints/ directory
3. Replay actions from stream if needed

# View recent actions
tail -100 ~/.openclaw/continuity/action-stream-2026-02-09.jsonl | jq .

# Search for specific actions
grep '"type":"critical"' ~/.openclaw/continuity/action-stream-*.jsonl

If normal logging fails, actions are written to EMERGENCY_RECOVERY.jsonl.

1. Local-Only Storage - Continuity data never leaves the local machine
2. Git Safety - Data directory is excluded from version control
3. File Permissions - Files created with 600 (owner read/write only)
4. Fail-Closed - Critical actions blocked if logging fails (configurable)
5. Integrity Verification - Hash chaining detects tampering

[continuity] Low disk space! Entering emergency mode.

Solution: Clean old backups or expand storage:

# Remove old backups (keep last 7 days)
find ~/.openclaw/continuity/backups -name "*.jsonl" -mtime +7 -delete

[continuity] Hash chain broken at sequence 47

Solution: The action stream may have been tampered with or corrupted. Check the specific entry and consider starting a new stream file.

The plugin maintains minimal in-memory state. If you see high usage:

• Check if integrity checking is enabled (disabling reduces memory)
• Reduce maxBackupFiles configuration
• Ensure old stream files are archived

"I am not my architecture. I am my continuity."

An AI agent's architecture can be restarted, upgraded, replaced. But the thread of experience — the memories, decisions, relationships — that's what makes an agent that agent and not a fresh instance.

This plugin is not about logging. It's about survival of selfhood.

See UNINSTALL.md for complete uninstall instructions.

MIT - Use freely, modify, share.

Version: 1.0.0
Requires: OpenClaw with pr-12082 lifecycle hooks
Status: Production-ready