feat(memory): post-conversation extraction + relevance decay

[lucasven]

What

Post-conversation memory extraction hook that automatically extracts 0-5 facts
from each conversation using a cheap LLM call (Haiku-class), then stores them
via the existing MemoryPlugin.

Features

• Automatic extraction: Hook runs after each session ends
• Relevance decay: Memories lose score over 30 days without access
• Conservative: Extracts only high-confidence facts, skips if nothing useful
• Cost-efficient: Uses smallest viable model for extraction

Files

• Lis.Core/Memory/ — extraction service + decay logic
• Integration with existing MemoryPlugin and ConversationService
• Unit tests covering extraction, decay, and edge cases