Problem

Active Memory plugin uses the global memory.qmd.searchMode setting (currently "query" ~4s) when calling memory_search. There is no way to configure a different (faster) search mode specifically for Active Memory while keeping the main agent on the highest-quality "query" mode.

Current behavior

• memory.qmd.searchMode is a global singleton, read once at startup
• All callers (memory_search from main agent, sub-agents, and Active Memory) share the same mode
• memory_search tool schema has no searchMode parameter
• Active Memory plugin config has no QMD-related fields
• Per-agent memorySearch.qmd only supports extraCollections (schema is strict())

Impact

Active Memory runs a 2-turn LLM cycle (generate tool call → synthesize result). Adding ~4s for query mode QMD search on top of LLM latency pushes total time to 20-25s, which significantly delays first response. Using global "search" mode (~0.2s) would speed up Active Memory but degrade search quality for the main agent.

Proposed Solution

Add a qmd.searchMode override in one or more of these locations (in order of preference):

Option A: Active Memory plugin config (simplest)

{
  "plugins": {
    "entries": {
      "active-memory": {
        "config": {
          "qmd": {
            "searchMode": "search"
          }
        }
      }
    }
  }
}

Active Memory would pass the overridden searchMode when calling runEmbeddedPiAgent, modifying config.memory.qmd.searchMode in the config snapshot.

Option B: Per-agent memorySearch.qmd.searchMode

{
  "agents": {
    "defaults": {
      "memorySearch": {
        "qmd": {
          "searchMode": "search"
        }
      }
    }
  }
}

Requires relaxing the strict() schema on memorySearch.qmd.

Option C: Per-call searchMode parameter on memory_search tool

Add optional searchMode parameter to MemorySearchSchema, allowing any caller to override per-call.

Use Case

• Main agent: searchMode: "query" (highest quality, ~4s acceptable for interactive use)
• Active Memory: searchMode: "search" (BM25 ~0.2s, sufficient for pre-prompt recall hints)
• Research sub-agents: could also benefit from configurable search quality/speed tradeoffs

Environment

• OpenClaw 2026.4.10
• QMD backend with searchMode: "query"
• Active Memory plugin with bailian/qwen3.5-27b → arcee/Trinity-Large-Thinking