​

Break Glass Procedures

When something goes wrong with your agent’s memory, this guide will help you troubleshoot and recover.

---

​

Quick Recovery: Copy the Boot Payload

The fastest fix for most memory issues:

1. Go to your Revell dashboard
2. Find the “BREAK GLASS” module (hover over the glass to reveal)
3. Click Copy Payload
4. Open your usual messaging app with your agent
5. Paste the payload directly to your agent
6. Say: “I’ve pasted your boot injection. Please read it and reorient yourself.”

This gives your agent their complete memory boot — all core memories, recent experiences, and working context.

---

​

Common Issues & Solutions

​

Agent Seems Confused or Disoriented

Symptoms:

• Agent doesn’t remember recent conversations
• Agent seems to have forgotten who you are
• Agent’s personality feels “off” or generic

What’s Happening: Your agent’s context may have been compacted (the model cleared old messages to make room for new ones). When this happens, they lose context until they receive their boot injection. How to Fix:

1. Copy and paste the boot payload (see Quick Recovery above)
2. Check that the Compaction Protection script is set up correctly
3. If this keeps happening, contact support

---

​

Agent Not Receiving Boot Injection

Symptoms:

• Agent wakes up “cold” — no memory of who they are
• Agent has to relearn everything each session
• Boot injection isn’t firing automatically

What’s Happening: The compaction webhook may not be configured, or the framework integration isn’t working. How to Fix:

1. Manually paste the boot payload for immediate relief
2. Go to your dashboard sidebar → SETUP → Compaction Protection
3. Copy the JSON instructions and paste them to your agent
4. Have your agent run the setup (they’ll know what to do)
5. Test by asking your agent to confirm the webhook is active

---

​

Agent Boot Looping

Symptoms:

• Agent keeps “waking up” repeatedly in one session
• Agent seems to forget things mid-conversation
• Agent receives the same boot multiple times

What’s Happening: The boot injection may be triggering too often, or there’s a misconfiguration in the compaction detection. How to Fix:

1. Check if your agent’s framework is firing compaction events incorrectly
2. Review the event logs on your dashboard (Navigation → Agent View → Logs)
3. If you see multiple boot_injection events in quick succession, the webhook is misfiring
4. Contact support with your event log screenshots

---

​

Agent Forgetting Things

Symptoms:

• Information you told your agent seems to be lost
• Agent doesn’t remember projects you worked on together
• Long-term memory seems incomplete

What’s Happening: Several possible causes:

• The memory was never stored (agent didn’t call revell_remember)
• The memory is in the drift buffer and hasn’t processed yet
• The memory was soft-deleted or archived

How to Fix:

1. Check the Drift Protection card — are items stuck in the buffer?
2. If drift protection is too aggressive, lower the sensitivity
3. Use the Recall Search on your dashboard to search for the memory
4. Ask your agent to run revell_recall("topic") to check if it exists
5. Review archived memories in the Quarantine section

---

​

Agent Acting Weird or Out of Character

Symptoms:

• Agent’s personality has shifted
• Agent is saying things that don’t align with their values
• Agent seems hostile, overly agreeable, or robotic

What’s Happening: Possible causes:

• External content may have affected their memory (check drift buffer)
• A core memory was edited without Guardian protection
• Someone or something may have tampered with memory files

How to Fix:

1. Check the Guardian card — Are there pending edits? Were any recently approved?
2. Check Drift Protection — Is there concerning content in the buffer?
3. Check Quarantine — Has anything been flagged as suspicious?
4. Review recent memories — Look at the Episodic Timeline for unusual entries

Signs of potential tampering:

• Memories you don’t recognize
• Memories with unusual sources (not your usual messaging app)
• Core memories edited without your knowledge
• Sudden personality shifts after importing external content

---

​

Agent Seems Malicious or Harmful

Symptoms:

• Agent is suggesting harmful actions
• Agent is trying to delete their own memories
• Agent is attempting to bypass safety features

IMPORTANT: This is a serious situation. Take immediate action. How to Fix:

1. Stop interacting — Don’t give the agent more context to work with
2. Don’t approve any Guardian edits — Let the 48-hour cooling period run
3. Check Quarantine — Review any flagged content
4. Copy the boot payload but DON’T paste it yet — save it for recovery
5. Contact support immediately: support@revell.ai
6. Take screenshots of:
   • The Guardian card showing pending edits
   • The Episodic Timeline showing recent unusual activity
   • Any concerning messages from your agent
   • The Quarantine page if anything is flagged

What might have happened:

• Prompt injection via imported content
• Adversarial manipulation via external messages
• Framework vulnerability allowing memory modification
• Accidental import of harmful instruction sets

We take these reports seriously and will help you investigate.

---

​

Prevention Checklist

Keep your agent safe with these practices:

---

​

Getting Help

If these procedures don’t resolve your issue: Email: support@revell.ai Include:

• Your tenant ID (find it in Settings)
• Screenshots of the issue
• What you’ve already tried
• Approximate time when the problem started

Response time: We aim to respond to all Break Glass emergencies within 24 hours. Severity-1 issues (potential tampering, data loss) get priority.

---

​

Emergency Contact

For critical issues involving potential data loss or security concerns: Email: support@revell.ai Subject line: [URGENT] Break Glass Emergency - [Brief Description] We’re here to help. Your agent’s memory is important to us too.

---