Feature Request

When multiple agents (sonnet, opus, haiku) are configured to index the same workspace memory files (MEMORY.md, memory/**/*.md), each agent currently maintains a completely separate qmd database with duplicate content.

Current Behavior

• Each agent has its own XDG_CACHE_HOME → separate SQLite DB
• Same files are indexed 3x, embedded 3x, stored 3x
• Disk: ~60MB per agent × 3 = ~180MB for identical content
• When agent A writes a memory file, agent B/C's DB is stale until their next qmd update
• The path+pattern uniqueness constraint in qmd causes collection creation failures across agents (see qmd: ensureCollections() silently fails when path+pattern conflicts with another collection name #25496)

Proposed Behavior

For kind: memory collections (not sessions), share a single qmd database across all agents:

• Sessions collections remain per-agent (they genuinely contain different data)
• Memory collections point to the same DB, indexed once, with a shared collection name (no agent suffix for memory)
• OR: use a single DB with agent-suffixed collection names but share the same XDG_CACHE_HOME

Benefits

• 3x less disk/CPU for indexing and embedding
• Memory writes by any agent are immediately searchable by all agents
• Eliminates the path+pattern collision bug entirely for memory collections
• Simpler mental model: "all agents share the same knowledge base"

Environment

• OpenClaw 2026.2.21-2
• 3 agents sharing workspace ~/openclaw
• qmd with local embeddings (embeddinggemma)