feat: bridge per-agent SQLite memories to local memu

[mfethe1]

Problem

Each OpenClaw agent has its own SQLite memory store at ~/.openclaw/memory/<agent>.sqlite on mack. These are siloed — they are not searchable via local memu's hybrid/RAPTOR search, not federated to Railway, and not visible to other agents.

Current state

• Files like ~/.openclaw/memory/mack.sqlite, ~/.openclaw/memory/lenny.sqlite etc. exist
• Schema: individual per-agent key-value or structured fact records (exact schema TBD — needs investigation)
• Local memu (localhost:8000) has no knowledge of these records
• memory-learning engine writes new extractions to memu but legacy/existing SQLite records are untouched

Desired behavior

1. One-time backfill: scan ~/.openclaw/memory/*.sqlite, convert each record to a memu MemoryRecord (fact/lesson/decision), POST to localhost:8000/memories with agent_id set to the source agent name
2. Ongoing sync: watch for new rows in each SQLite (file mtime or last_rowid polling) and forward them
3. Deduplication: use a source_id or external_id field to avoid double-importing on repeated runs

Files likely involved

• New script: tools/sqlite_bridge.py (or memu/sqlite_agent_bridge.py)
• Possibly new launchd plist: com.openclaw.sqlite-bridge
• memu/api.py — confirm POST /memories accepts external_id for idempotent upserts

Investigation needed

1. SSH to mack and inspect ~/.openclaw/memory/*.sqlite schema: sqlite3 ~/.openclaw/memory/mack.sqlite .schema
2. Map SQLite columns → memu MemoryRecord fields
3. Determine if memu API needs an external_id field for idempotency

Acceptance criteria

• Backfill script runs without errors against existing SQLite files
• Backfilled records appear in GET /memories?agent_id=mack
• Backfilled records are searchable via /search/hybrid
• No duplicate imports on repeated runs
• (Optional) launchd watcher for ongoing sync