Benchmark: build realistic test corpus from agent memory patterns

[bm-clawd]

Part of #10

Build a realistic test corpus that mirrors how agents actually use memory in production.

Requirements

• Anonymized markdown files covering all the patterns agents use:
  • MEMORY.md — curated long-term facts, preferences, decisions
  • memory/YYYY-MM-DD.md — daily notes with events, checks, conversations
  • memory/tasks/*.md — structured tasks with schema fields
  • memory/topics/*.md — research notes, competitive analysis
  • memory/people/*.md — person notes with relations
• Realistic size: ~30-50 files, ~50-100KB total
• Must exercise all query categories: exact facts, semantic, temporal, relational, cross-note, needle-in-haystack
• Include evolving facts (same topic updated across multiple daily notes)
• Include BM-specific features: observations [key] value, relations relates_to [[Entity]], frontmatter schemas

Deliverable

• benchmark/corpus/ with all files
• benchmark/queries.json with 50+ annotated queries
• Ground truth: which files/chunks contain the answer for each query