Summary

QMD backend falls back to builtin and never recovers, even after gateway restart and increasing timeout.

Environment

• OpenClaw version: 2026.2.3-1
• OS: macOS (arm64)
• QMD: installed via bun (/Users/jmworks/.bun/bin/qmd)

Config

"memory": {
  "backend": "qmd",
  "qmd": {
    "command": "/Users/jmworks/.bun/bin/qmd",
    "limits": {
      "timeoutMs": 120000
    }
  }
}

Steps to Reproduce

1. Configure memory.backend = "qmd"
2. First memory_search call triggers QMD query
3. QMD times out (downloading/loading models takes ~3 minutes on first run)
4. OpenClaw logs: [memory] qmd memory failed; switching to builtin index
5. After timeout extended to 120s and gateway restarted, QMD is never called again
6. All subsequent memory_search calls immediately fallback to builtin (which then fails due to no API key)

Expected Behavior

After gateway restart with longer timeout and models pre-downloaded, QMD should be attempted again.

Actual Behavior

Once QMD fails, OpenClaw seems to permanently skip QMD and go straight to builtin fallback, even after:

• Increasing timeoutMs from 4000 → 30000 → 120000
• Copying models to correct XDG cache location
• Multiple gateway restarts (SIGUSR1 and full process restart)

Logs

2026-02-05T15:11:04.707Z [memory] qmd query failed: Error: qmd query ... timed out after 4000ms
2026-02-05T15:11:04.707Z [memory] qmd memory failed; switching to builtin index
...
2026-02-05T15:29:20.473Z [memory] qmd query failed: Error: qmd query ... timed out after 30000ms
2026-02-05T15:29:20.474Z [memory] qmd memory failed; switching to builtin index

After config change to 120000ms and restart, no new QMD-related logs appear - it seems to skip QMD entirely.

Workaround Attempted

• Pre-warming QMD manually with same XDG env vars works fine
• Models downloaded and copied to ~/.openclaw/agents/main/qmd/xdg-cache/qmd/models/
• QMD direct execution returns results in ~10 seconds after warm-up