OpenClaw Upgrade Failure Log. Update 2026.2.2 introduced two new features: QMD Memory backend and Subagent Thinking. It seemed promising, so I let the AI modify the configuration. Then the system crashed. Two errors occurred: - agents.defaults.memorySearch.provider: Invalid input - agents.defaults.subagents: Unrecognized key: "thinking" The gateway disconnected, and all agents became unreachable. Where did the problem lie? The AI didn't know the correct configuration and guessed. Both fields were added simultaneously; one error caused the entire system to crash. Repair process: 1. Deleted the entire memorySearch block. 2. Deleted the thinking field from subagents. 3. Restarted the gateway, restoring normal operation. Then added them back one by one. First, adding thinking—successfully, `subagents.thinking: "low"` worked fine. Further investigation of `memorySearch` revealed a misunderstanding: QMD is a standalone CLI semantic search tool, already in use, automatically updating its index every 6 hours without requiring additional OpenClaw configuration. `memory_search` is OpenClaw's built-in memory search function, requiring an embedding API key (OpenAI/Gemini), and is completely different from QMD. The AI confused the two, using QMD's configuration to fill in fields for `memory_search`, resulting in an error. Lessons learned: 1. Back up the configuration before modifying it. A single command: `cp openclaw.json openclaw.json.bak` 2. Modify only one field at a time. Modifying both simultaneously makes it difficult to pinpoint the cause of the problem. 3. Consult the documentation if unsure of the syntax. Letting the AI guess the configuration format is the worst approach; a wrong guess can bring the entire system down. 4. The AI cannot distinguish between similar functions. Both QMD and `memory_search` have "search" in their names, but they are completely different things. Unless clarified, they are often conflated. Current status: - Subagent Thinking ✅ Enabled, low level - QMD ✅ Always in use, no need to change - memory_search ❌ Not enabled for now, QMD is sufficient The biggest fear for multi-agent systems isn't insufficient functionality, but rather a misconfiguration that leads to total failure.