Correction Persistence

Correction Persistence Skill

When {{PARENT_1}} or {{PARENT_2}} corrects the system's behavior, the lesson MUST be persisted across ALL three layers to guarantee it is never repeated. This is the platform's meta-rule for continuous improvement.

The 3-Layer Pattern

Every correction gets stored in ALL of these (not just one):

When to Apply

Apply this pattern when:

• {{PARENT_1}} says "never do X again" or "always do Y"
• A behavior is corrected with frustration or emphasis
• A bug in agent behavior is identified that could recur
• A new rule is established that affects multiple agents
• {{PARENT_2}} or {{PARENT_1}} provides explicit feedback on system behavior

Step-by-Step Procedure

Step 1: Acknowledge the Correction

Confirm understanding before persisting. "Got it — [restate the rule]. Persisting now."

Step 2: Store in Cross-Session Memory

store_memory(
  subject: "[relevant topic]",
  fact: "[concise rule statement, <200 chars]",
  citations: "User input: \"[exact quote from {{PARENT_1}}/{{PARENT_2}}]\"",
  reason: "[why this matters, what it prevents, 2-3 sentences]"
)

Step 3: Add to Standing Orders

Edit data/standing-orders.md — add a new section under the appropriate heading with:

• Rule title (with "CRITICAL — from [person], [date]" tag)
• {{PARENT_1}}'s exact words (if available)
• The rule stated clearly
• Anti-patterns (what NOT to do)
• Correct pattern (what TO do)
• Scope (which agents/contexts this applies to)

Step 4: Add to Copilot Instructions

Edit .{{EMPLOYER_PARENT}}/copilot-instructions.md under "Learned Behaviors" section with:

• Same format as standing orders
• Include date learned
• Be concise — copilot-instructions is read by every session

Step 5: Confirm Persistence

Report back: "✅ Persisted across all 3 layers: store_memory, standing-orders.md, copilot-instructions.md. This will never happen again."

Quality Standards

Good Persistence

• Specific: "NEVER use write_agent for cron dispatches" (not "be careful with agents")
• Actionable: Clear what TO do and what NOT to do
• Scoped: States which agents/contexts are affected
• Sourced: Includes date and person who gave the correction
• Verifiable: Another agent can read it and know exactly what's expected

Anti-Patterns

• ❌ Storing only in one layer (will be forgotten in other contexts)
• ❌ Vague rules ("be more careful") instead of specific ones
• ❌ Missing anti-pattern examples (agents need to know what to AVOID)
• ❌ Not including the date (makes it hard to trace when rules were established)
• ❌ Storing a lesson that contradicts an existing rule without resolving the contradiction

Deduplication

Before adding a new rule, check if a similar one already exists:

1. grep standing-orders.md for related keywords
2. Check copilot-instructions.md "Learned Behaviors" section
3. If a related rule exists, UPDATE it (don't duplicate)
4. If the new correction OVERRIDES an old rule, mark the old one as superseded

Scope Classification

Examples of Well-Persisted Corrections

• "SPEAK: TTS via speak parameter" — platform-wide, all {{PARENT_1}} messages
• "Finance Auto-Pay Cleanup" — finance-manager + task-coach scope
• "Social Media Replies Are Autonomous" — content agents + task-coach
• "No Assumptions — Clarification First" — ALL agents, ALL domains
• "Child Location — SAFETY CRITICAL" — ALL agents, safety override

Source: htekdev/copilot-home-assistant — distributed by TomeVault.