MEMORY FIRST (MCP Memory Server)

• BEFORE starting work: Search existing memory

  • Search matches: entity names, entity types, AND observation content
  • Best practice: Start with specific terms (foggysky, foggysky-authentication)
  • Multi-term searches work if terms appear in observations
  • Use hyphenated entity names for precise retrieval

• AFTER completing tasks: Add information with UTC timestamp

  • Format: YYYY-MM-DDTHH:mm:ssZ [observation]
  • Timestamp by environment:
    • Claude Desktop: new Date().toISOString() in analysis tool
    • Claude Code: date -u +"%Y-%m-%dT%H:%M:%SZ"
    • Fallback: State timestamp may be inaccurate
  • For contradictions: Remove the old memory

• Entity Structure (Fine-Grained):

  • ✅ ONE fact per observation - never combine multiple facts
  • ✅ Atomic entities: Split by component/feature/concern
  • ✅ Naming: project component feature keywords (use searchable full words)
  • ❌ No code blocks, paragraphs, or documentation dumps

• Use Relations:

  • Link entities with depends-on, solves, uses, implements
  • Avoid duplicating information across entities