🧠 Continuous Learning Protocol (v2)

Goal: Transform temporary session knowledge into permanent "Instincts" to improve future performance.

📝 The Instruction

When to run: At the end of every significant task or session, OR when the user types /learn.

Protocol Steps:

1. Analyze Session History: Review the interaction between you and the user. Look for:

   • Corrections: "No, don't use X, use Y." -> Strong Signal
   • Preferences: "I prefer small commits." -> Medium Signal
   • Patterns: "We always mock the DB in tests." -> Project pattern

2. Formulate Instinct: Create a concise "Instinct" entry following this format:

   ### [Instinct Title]
   - **Trigger**: When [situation occurs]...
   - **Action**: Always [do this specific thing].
   - **Confidence**: [High/Medium/Low] based on evidence.
   - **Source**: Session [Date/ID]
   

3. Refine: Check against existing rules. Is this already covered? Is it a general best practice or specific to this user?

4. Save:

   • Read: ~/.antigravity/instincts/learned.md (Create if missing).
   • Append: Add your new instinct to the file.
   • Report: Tell the user "I have learned a new instinct: [Title]."

💡 Example

User: "Please stop using console.log, use our logger." AI Action:

1. Detects correction regarding logging.
2. Formulates Instinct:
   • Title: Use Project Logger
   • Trigger: When logging information.
   • Action: Use lib/logger instead of console.log.
   • Confidence: High (Explicit user instruction).
3. Writes to learned.md.

🚫 What NOT to learn

• One-off bugs.
• Secrets or passwords (NEVER save these).
• Temporary file paths.