Install
openclaw skills install z1-matrix-memory-palaceBuild and operate a file-driven Memory Palace for multi-agent systems, combining a spatial memory shell with a continuously maintained LLM Wiki reflection layer. Use when designing or deploying long-term memory architecture for agent teams, especially for (1) project rooms + task corridors + reflection wings, (2) low-token file-based coordination buses, (3) converting work artifacts into principles, prompt kernels, failure patterns, and thinking paths, (4) building a silent librarian/archivist agent, or (5) migrating scattered protocol/history files into a structured memory operating system. Includes source attribution and fusion guidance from Karpathy's llm-wiki idea and Jeff Pierce's memory-palace approach.
openclaw skills install z1-matrix-memory-palace以“空间骨架 + 编译灵魂 + 低 token 调度宪法”为核心,构建长期运行的多 Agent 记忆操作系统。
将长期记忆拆成三层:
一句话总纲: 用 Memory Palace 组织空间,用 LLM Wiki 编译知识,用文件驱动宪法压低噪声。
必须明确标注两条来源:
核心启发:
一句话: 知识不是临时召回,而是持续编译。
核心启发:
一句话: 记忆不属于模型,记忆属于系统。
本技能不照抄任何单一方案,而是融合为:
适用于:
不适用于:
全局总控台,负责:
按 Agent 分房:
按项目分房,例如:
任务运行态总线:
长期编译区:
低频历史归档层。
任务状态优先由 task card 和目录位置承载,而不是由聊天上下文承载。
能通过输入/输出路径锚定,就不要重复灌输背景。
同一任务默认一条主执行链,不开多子对话分叉。
专设 Archivist / Librarian,只在后台读取 completed cards 与 project rooms,编译长期记忆,不参与前线对话。
当 completed card 出现后,判断是否编译:
对已有高价值协议:
必须同时满足:
若你需要后台记忆编译员:
优先采用:
YYYY-MM-DD_<slug>.md<scope>-<slug>-<date><project_slug>/<semantic_slug>_<YYYY-MM-DD>.mdz1_<protocol_name>_v<major>_<YYYY-MM-DD>.mdRUNBOOK_V<major>_<YYYY-MM-DD>.mdreferences/architecture.mdreferences/publishing-note.mdreferences/naming-quickref.mdreferences/release-copy-zh.mdreferences/release-copy-en.md当使用本技能帮助他人落地时,最终交付应至少包含:
Z1矩阵宫殿记忆法,不是做一个会搜的库,而是做一个能持续沉淀、持续编译、持续收束噪声的多 Agent 记忆中枢。