OpenClaw Agent Optimization Guide

QMD Memory Backend, Thinking Mode, LCM, & Memory Search Tuning — Updated March 2026
Prepared by Mateo Vargas, BrightBlitz Marketing

📋 Overview — What We Changed and Why

🧠 Change 1: Thinking Mode → HIGH

What Thinking Mode Does

Anthropic's Claude models support "extended thinking" — an internal reasoning step before generating a response. At low, the model gives quick answers. At high, it reasons through the problem step by step before responding. The difference is dramatic for complex tasks.

Real-World Example

After switching to HIGH, our agent caught its own mistakes from 10 minutes earlier in the same conversation. Same model, same context — just actually reasoning now.

Available Levels

Config

{
  "agents": {
    "defaults": {
      "thinkingDefault": "high",
      "subagents": {
        "thinking": "high"
      }
    }
  }
}

💾 Change 2: QMD Memory Backend

What QMD Is

QMD (Query Markup Documents) is a local hybrid search engine by Tobi Lütke (Shopify CEO). It replaces OpenClaw's default search with three layers:

1. BM25 Keyword Matching — finds exact terms (IP addresses, error codes, names, plugin versions)
2. Vector Semantic Search — finds related content even when different words are used
3. LLM Re-ranking — a small local model confirms which results are actually relevant

Everything runs locally. No API keys needed for the search itself. ~2GB of GGUF models auto-download on first use.

Installation

# Install QMD globally (npm or bun)
npm install -g @tobilu/qmd
# or: bun install -g @tobilu/qmd

# Verify
qmd --version  # Should show 2.0.1+

Config

{
  "memory": {
    "backend": "qmd"
  }
}

🔧 Change 3: QMD Memory Search Tuning

QMD works out of the box, but three settings that are OFF by default make a significant difference:

3A. MMR (Maximal Marginal Relevance) — Diversity Re-ranking

What MMR does: After finding the top results, it re-ranks them to maximize diversity. If result #2 is too similar to result #1, it gets demoted in favor of a result that adds new information.

The lambda parameter (0.0 to 1.0):

• 0.0 = maximum diversity (ignore relevance, just pick different results)
• 1.0 = maximum relevance (ignore diversity, return the best matches even if redundant)
• 0.7 = recommended balance — prioritize relevance but actively suppress duplicates

3B. Temporal Decay — Freshness Bias

How it works: Applies an exponential decay to result scores based on age. With a 30-day half-life:

Old notes still appear if they're the only match — they just don't outrank fresh notes on the same topic.

3C. Embedding Cache — Cost Optimization

Combined Config

{
  "agents": {
    "defaults": {
      "memorySearch": {
        "cache": {
          "enabled": true,
          "maxEntries": 50000
        },
        "query": {
          "hybrid": {
            "enabled": true,
            "vectorWeight": 0.7,
            "textWeight": 0.3,
            "candidateMultiplier": 4,
            "mmr": {
              "enabled": true,
              "lambda": 0.7
            },
            "temporalDecay": {
              "enabled": true,
              "halfLifeDays": 30
            }
          }
        }
      }
    }
  }
}

📝 Change 4: Session Transcript Indexing

What This Does

QMD indexes past session transcripts so memory_search can find things you discussed even if the agent didn't write them to a file. Set a retention window to control how far back it indexes.

Config

{
  "memory": {
    "qmd": {
      "sessions": {
        "enabled": true,
        "retentionDays": 30
      }
    }
  }
}

📌 Change 5: Memory Citations

Config

{
  "memory": {
    "citations": "auto"
  }
}

📂 Change 6: Extra Index Paths

Config

{
  "memory": {
    "qmd": {
      "includeDefaultMemory": true,
      "paths": [
        { "path": "reference/", "name": "reference" },
        { "path": "runbooks/", "name": "runbooks" },
        { "path": "sops/", "name": "sops" }
      ]
    }
  }
}

Add whatever folders contain persistent knowledge your agents should be able to search. Adjust the paths to match your workspace structure.

🔄 Change 7: LCM v0.4.0 — Session Scoping (March 2026)

LCM v0.4.0 adds session scoping — you can tell it which sessions to ignore entirely and which should be read-only.

Key New Settings

Config

{
  "plugins": {
    "entries": {
      "lossless-claw": {
        "enabled": true,
        "config": {
          "freshTailCount": 32,
          "incrementalMaxDepth": -1,
          "contextThreshold": 0.75,
          "ignoreSessionPatterns": [
            "agent:*:cron:**"
          ],
          "statelessSessionPatterns": [
            "agent:*:subagent:**"
          ],
          "skipStatelessSessions": true
        }
      }
    }
  }
}

Session Persistence

Also recommended: increase OpenClaw's session idle timeout so long conversations don't reset:

{
  "session": {
    "reset": {
      "mode": "idle",
      "idleMinutes": 10080
    }
  }
}

10080 = 7 days. Prevents session resets during weekends or quiet periods.

Heartbeat Pruning (Optional)

If you run heartbeats, HEARTBEAT_OK responses accumulate in LCM. Set this environment variable (not plugin config) to clean them:

LCM_PRUNE_HEARTBEAT_OK=true

Add via systemd override or your process manager's environment config.

📦 Change 8: QMD 2.0 Update (March 2026)

New Features Worth Knowing

• Intent parameter — Pass --intent "context" to disambiguate searches. "performance" with intent "server speed" finds different results than "performance" with intent "employee review."
• Collection ignore patterns — Exclude files from indexing: ignore: ["Sessions/**", "*.tmp"]
• Query --explain — Debug tool showing retrieval scores, RRF contributions, and reranker scores per result
• SDK / library mode — QMD can now be used as a Node.js library, not just CLI

Update Command

# npm:
npm install -g @tobilu/qmd@latest

# bun:
bun install -g @tobilu/qmd@latest

# Verify:
qmd --version  # Should show 2.0.1+

Collections Setup

QMD 2.0 still needs collections configured. If you haven't set them up, or after a fresh install:

# Index your memory files
qmd collection add ~/path-to/memory/ --name "daily-memory"

# Index deliverables, research, etc
qmd collection add ~/path-to/deliverables/ --name "deliverables"

# Generate embeddings (first time takes a few minutes on CPU)
qmd embed

📋 Full Config Block (Copy-Paste Ready)

{
  "memory": {
    "backend": "qmd",
    "citations": "auto",
    "qmd": {
      "includeDefaultMemory": true,
      "paths": [
        { "path": "reference/", "name": "reference" },
        { "path": "runbooks/", "name": "runbooks" },
        { "path": "sops/", "name": "sops" }
      ],
      "sessions": {
        "enabled": true,
        "retentionDays": 30
      },
      "update": {
        "interval": "5m",
        "debounceMs": 15000
      },
      "limits": {
        "maxResults": 8,
        "timeoutMs": 6000
      }
    }
  },
  "session": {
    "reset": {
      "mode": "idle",
      "idleMinutes": 10080
    }
  },
  "plugins": {
    "entries": {
      "lossless-claw": {
        "enabled": true,
        "config": {
          "freshTailCount": 32,
          "incrementalMaxDepth": -1,
          "contextThreshold": 0.75,
          "ignoreSessionPatterns": ["agent:*:cron:**"],
          "statelessSessionPatterns": ["agent:*:subagent:**"],
          "skipStatelessSessions": true
        }
      }
    }
  },
  "agents": {
    "defaults": {
      "thinkingDefault": "high",
      "subagents": {
        "thinking": "high"
      },
      "memorySearch": {
        "cache": {
          "enabled": true,
          "maxEntries": 50000
        },
        "query": {
          "hybrid": {
            "enabled": true,
            "vectorWeight": 0.7,
            "textWeight": 0.3,
            "candidateMultiplier": 4,
            "mmr": {
              "enabled": true,
              "lambda": 0.7
            },
            "temporalDecay": {
              "enabled": true,
              "halfLifeDays": 30
            }
          }
        }
      }
    }
  }
}

🛠️ Installation Steps

bun install -g https://github.com/tobi/qmd

# Via bot:
# Paste the config block and ask your agent to apply it via config.patch

# Or manually:
nano ~/.openclaw/openclaw.json
# Merge the config, save, then:
openclaw gateway restart

⚠️ Gotchas & Lessons Learned

✅ Verification Checklist

Prepared by Mateo Vargas — BrightBlitz Marketing
Originally February 27, 2026 — Updated March 21, 2026