OpenClawの記憶問題を解決する方法

I Fixed OpenClaw's Biggest Problem (Memory) OpenClaw's default memory system causes AI agents to lose context between sessions. Here's how I rewired the entire system so it actually works. Core Problem Agents experience "amnesia" because short-term memory fills quickly with tool outputs and long messages. When compaction occurs, important context disappears if it wasn't saved to long-term memory files first. Key Solutions File Organization: Create structured memory folders: memory/projects/, memory/decisions/, and memory/lessons/. This separates permanent facts, active projects, decisions, and lessons learned rather than dumping everything into one file. Compaction Settings: Specific configuration changes preserve recent context, including setting keepRecentTokens to 20000 and recentTurnsPreserve to 4 to keep the last four user/assistant exchanges intact without summarization. Memory Flush Customization: A custom prompt tells the agent exactly what to save — "decisions made, corrections given, constraints or preferences" — rather than using generic defaults. Context Pruning: Enabling this feature automatically trims bloated tool outputs before they consume the context window. QMD Installation: This tool enables semantic search across conversation transcripts, serving as a safety net for unsaved context. Behavioral Habits: - Explicitly instruct agents to save important decisions - End sessions by listing key points to preserve - Review and delete irrelevant memory files weekly The result: agents progressed from forgetting names between sessions to remembering conversations from weeks prior.