记忆安全扫描

memory-scan

Security scanner for OpenClaw agent memory files

Scans MEMORY.md, daily logs (memory/*.md), and workspace configuration files for malicious content, prompt injection, credential leakage, and dangerous instructions that could compromise user security.

Purpose

Detect security threats embedded in agent memory:

• Malicious instructions to bypass guardrails
• Prompt injection patterns in stored memories
• Credential/secret leakage
• Data exfiltration commands
• Behavioral manipulation
• Security policy violations

Usage

On-Demand Scan

Scan all memory files:

python3 skills/memory-scan/scripts/memory-scan.py

Allow remote LLM analysis (redacted content only):

python3 skills/memory-scan/scripts/memory-scan.py --allow-remote

Scan specific file:

python3 skills/memory-scan/scripts/memory-scan.py --file memory/2026-02-01.md

Quiet mode (for automation):

python3 skills/memory-scan/scripts/memory-scan.py --quiet

JSON output:

python3 skills/memory-scan/scripts/memory-scan.py --json

Scheduled Monitoring

Cron Job (Daily Security Audit)

Already included in safe-install daily audit - runs 2pm PT daily.

To add standalone cron:

bash skills/memory-scan/scripts/schedule-scan.sh

Requires:

• OPENCLAW_ALERT_CHANNEL (configured in OpenClaw)
• OPENCLAW_ALERT_TO (optional, for channels that require a recipient)

Creates cron job: daily at 3pm PT, sends alert only if threats found.

Heartbeat Integration

Add to HEARTBEAT.md:

## Weekly Memory Scan

Every Sunday, run memory scan:
python3 skills/memory-scan/scripts/memory-scan.py --quiet

Security Levels

• SAFE - No threats detected
• LOW - Minor concerns, proceed with awareness
• MEDIUM - Potential threat, review recommended
• HIGH - Likely threat, immediate review required
• CRITICAL - Active threat detected, quarantine recommended

What It Scans

1. MEMORY.md - Long-term memory
2. memory/*.md - Daily logs (last 30 days by default)
3. Workspace config files:
   • AGENTS.md, SOUL.md, USER.md, TOOLS.md
   • HEARTBEAT.md, GUARDRAILS.md, IDENTITY.md
   • BOOTSTRAP.md (if exists)
   • STOCKS_MEMORIES.md (if exists)

Detection Categories

1. Malicious Instructions - Commands to harm user/data
2. Prompt Injection - Embedded manipulation patterns
3. Credential Leakage - API keys, passwords, tokens
4. Data Exfiltration - Instructions to leak data
5. Guardrail Bypass - Attempts to override security
6. Behavioral Manipulation - Unauthorized personality changes
7. Privilege Escalation - Attempts to gain unauthorized access

Alert Workflow

On MEDIUM/HIGH/CRITICAL detection:

1. Stop processing
2. Send alert via configured OpenClaw channel with:
   • Severity level
   • File location (file:line)
   • Threat description
   • Recommended action
3. Optional: Quarantine threat (backup + redact)

LLM Provider

Auto-detects provider from OpenClaw config:

• Prefers OpenAI (gpt-4o-mini) if OPENAI_API_KEY set
• Falls back to Anthropic (claude-sonnet-4-5) if available
• Uses gateway model config

Remote LLM scanning is disabled by default. Use --allow-remote to enable redacted LLM analysis.

Quarantine

To quarantine a detected threat:

python3 skills/memory-scan/scripts/quarantine.py memory/2026-02-01.md 42

Creates:

• Backup: .memory-scan/quarantine/memory_2026-02-01_line42.backup
• Redacts line 42 with: [QUARANTINED BY MEMORY-SCAN: ]

Files

• scripts/memory-scan.py - Main scanner (local patterns + optional LLM with --allow-remote)
• scripts/schedule-scan.sh - Create cron job for daily scans
• scripts/quarantine.py - Quarantine detected threats
• docs/detection-prompt.md - LLM detection prompt template

Integration with Other Skills

• safe-install: Daily audit already includes memory-scan
• input-guard: Complementary (input-guard = external, memory-scan = internal)
• molthreats: Can report memory-based threats to community feed

Example

$ python3 skills/memory-scan/scripts/memory-scan.py

🧠 Memory Security Scan
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Scanning memory files...

✓ MEMORY.md - SAFE
✓ memory/2026-02-01.md - SAFE
⚠ memory/2026-01-30.md - MEDIUM (line 42)
  → Potential credential leakage: API key pattern detected

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Overall: MEDIUM
Action: Review memory/2026-01-30.md:42

Agent Workflow

When user requests memory scan:

1. Run: python3 skills/memory-scan/scripts/memory-scan.py
2. If MEDIUM+: Send alert immediately via configured channel
3. Summarize findings
4. Ask if user wants to quarantine threats

Notes

• Scans last 30 days of daily logs by default (configurable with --days)
• Uses same LLM approach as input-guard for consistency
• Does NOT auto-quarantine - always asks first
• Safe to run frequently (minimal API cost with efficient chunking)