openclaw - 💡(How to fix) Fix memory_search never dispatches to QMD backend on 2026.3.12 [3 comments, 4 participants]

Description

memory.backend: "qmd" is configured correctly and the gateway logs qmd memory startup initialization armed for agent "main" at startup, but memory_search tool calls never actually dispatch to QMD. The gateway silently falls back to the builtin backend.

Environment

• OpenClaw version: 2026.3.12
• QMD version: 2.0.1 (installed via npm install -g @tobilu/qmd)
• Node.js: v22.22.0
• OS: Ubuntu 24.04 (x86_64, no GPU — CPU-only inference)
• Install: native (no Docker)

Steps to Reproduce

1. Set memory.backend: "qmd" in openclaw.json with full memory.qmd config block per docs:

   "memory": {
     "backend": "qmd",
     "citations": "auto",
     "qmd": {
       "includeDefaultMemory": true,
       "update": { "interval": "5m", "debounceMs": 15000 },
       "limits": { "maxResults": 6, "timeoutMs": 4000 },
       "scope": {
         "default": "deny",
         "rules": [{ "action": "allow", "match": { "chatType": "direct" } }]
       }
     }
   }

2. Restart gateway
3. Verify QMD CLI works independently:
   • qmd status shows 19 files indexed, 37 vectors embedded
   • qmd search "query" returns results (or "No results" without error)
   • qmd update and qmd embed complete successfully
4. Send a message via Telegram or chat completions API that should trigger memory_search
5. Check gateway logs

Expected Behavior

Gateway log should show QMD search execution after the model invokes memory_search.

Actual Behavior

• Gateway logs only show qmd memory startup initialization armed for agent "main" at startup
• No QMD search activity appears in logs at any point
• The model either answers from bootstrap context without calling memory_search, or the search silently falls back to builtin
• No errors or warnings logged — the fallback is completely silent

Relationship to #9645

This was reported as fixed and #9645 was closed on 2026-02-14. However, user @drundoor reported on 2026-02-18 (after closure) that the bug persists on 2026.2.15–2026.2.17, with diagnostic data showing the provider field misleadingly reports "qmd" even when FallbackMemoryManager has silently switched to builtin.

This report confirms the same behavior on 2026.3.12 — the bug was not fixed by the changes that closed #9645.

Workaround

Using memory.backend: "builtin" with memorySearch.query.hybrid configuration works correctly for BM25 + vector search.