← Back to Skills Marketplace
z1one0415

Z1矩阵宫殿记忆法

by z1one0415 · GitHub ↗ · v1.0.1 · MIT-0
cross-platform ✓ Security Clean
133
Downloads
0
Stars
0
Active Installs
2
Versions
Install in OpenClaw
/install z1-matrix-memory-palace
Description
Build and operate a file-driven Memory Palace for multi-agent systems, combining a spatial memory shell with a continuously maintained LLM Wiki reflection la...
README (SKILL.md)

Z1矩阵宫殿记忆法

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。

核心定义

将长期记忆拆成三层:

  1. Raw Layer:工作区原始产物、协议、日志、任务卡
  2. Palace Layer:房间、走廊、项目室、总厅、索引、路由
  3. Wiki / Reflection Layer:原则页、提示词内核、失败模式、思考路径、候选宪法

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。

参考来源与融合思路

必须明确标注两条来源:

来源A:Karpathy 的 llm-wiki 思路

核心启发:

  • 不是每次查询都从原始材料重新拼装知识
  • 而是由 LLM 持续维护一个持久、结构化、可交叉引用的 wiki 中层
  • 新材料进入后,不只是索引,而是编译进既有知识结构
  • 问答结果也可回写为新页面

一句话: 知识不是临时召回,而是持续编译。

来源B:Jeff Pierce 的 memory-palace 思路

核心启发:

  • 记忆应独立于模型上下文存在
  • 记忆应具备跨实例、跨 Agent 的共享与检索能力
  • 记忆层适合承接关系、消息、分区、长期存储
  • 记忆系统应成为模型旁边的长期层,而不是模型内部的幻觉延长线

一句话: 记忆不属于模型,记忆属于系统。

融合方法

本技能不照抄任何单一方案,而是融合为:

  • Memory Palace 负责:空间分层、项目房间、任务走廊、总控路由
  • LLM Wiki 负责:持续编译原则、方法、失败模式、思考路径
  • 文件驱动宪法 负责:把通信从聊天迁移到路径、卡片、物理产物

适用对象

适用于:

  • 多 Agent 团队
  • 长期项目知识沉淀
  • 跨任务经验复用
  • 希望减少上下文漂移与 token 浪费的系统
  • 需要把“项目事实”与“长期原则”分层的知识架构

不适用于:

  • 一次性临时笔记
  • 只有单轮聊天、无长期维护需求的场景
  • 没有物理文件产出的纯对话式工作流

默认空间拓扑

1. Grand Hall

全局总控台,负责:

  • active 任务总览
  • completed 链路总览
  • reflection 覆盖总览
  • 核心协议入口

2. Chambers

按 Agent 分房:

  • 01_core
  • 02_intel
  • 03_pub
  • 04_code
  • 05_media
  • 可扩展 archivist 等后台角色

3. Project Rooms

按项目分房,例如:

  • raw_stone
  • karpathy_daily
  • lobster_legion

4. Dispatch Corridor

任务运行态总线:

  • active/
  • blocked/
  • completed/

5. Reflection Wing

长期编译区:

  • principles/
  • prompt_kernels/
  • failure_patterns/
  • thinking_paths/
  • constitution_candidates/

6. Archive Basement

低频历史归档层。

运行原则

原则1:路径即状态

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。

原则2:文件优先于对话

能通过输入/输出路径锚定,就不要重复灌输背景。

原则3:项目沉淀与原则提炼分离

  • project room 存项目事实
  • reflection wing 存长期逻辑
  • constitution candidate 只接收反复验证的硬规则

原则4:一题一主链

同一任务默认一条主执行链,不开多子对话分叉。

原则5:静默 Librarian

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。

标准工作流

A. 建骨架

  1. 创建 grand_hall / chambers / project_rooms / dispatch_corridor / reflection_wing
  2. 创建总控台、命名规范、索引与日志
  3. 明确每个房间的职责边界

B. 让任务上走廊

  1. 每条主链任务建立唯一 task card
  2. active / blocked / completed 三态迁移
  3. 卡片中必须写明输入/输出路径、验收、下一步

C. 让项目进房间

  1. 为长期项目建立 project room
  2. 写 index、chain_map、reflection_backfill
  3. 把 corridor 卡与 reflection 页面回链到项目房间

D. 让经验进反思侧翼

当 completed card 出现后,判断是否编译:

  • principle card
  • prompt kernel
  • failure pattern
  • thinking path
  • constitution candidate

E. 做旧记忆回填

对已有高价值协议:

  1. 保留原始文件作为历史源
  2. 在 palace 对应 chamber 或 project room 中建立镜像文件
  3. 建立 legacy_index
  4. 让 grand hall 和 reflection 可以导航到这些镜像

记忆代谢规则

只留在 Project Room 的内容

  • 单一项目事实
  • 样本与阶段痕迹
  • 不具跨任务复用价值的内容

进入 Reflection Wing 的内容

  • 多任务复现有效
  • 能压缩为原则 / kernel / failure / path
  • 能降低未来 token 成本或错误率

升级为 Constitution Candidate 的内容

必须同时满足:

  1. 多任务验证
  2. 偏离代价显著
  3. 属于长期默认行为
  4. 不是局部技巧

Librarian / Archivist 规则

若你需要后台记忆编译员:

  • 只读取已落盘文件
  • 只在 completed 后触发编译
  • 不直接对外交互
  • 不把聊天当主证据
  • 不把局部经验滥升为宪法

命名与路由规范

优先采用:

  • Corridor 卡片:YYYY-MM-DD_\x3Cslug>.md
  • task_id:\x3Cscope>-\x3Cslug>-\x3Cdate>
  • Project Room:\x3Cproject_slug>/
  • Reflection 页面:\x3Csemantic_slug>_\x3CYYYY-MM-DD>.md
  • Core Protocol:z1_\x3Cprotocol_name>_v\x3Cmajor>_\x3CYYYY-MM-DD>.md
  • Agent Runbook:RUNBOOK_V\x3Cmajor>_\x3CYYYY-MM-DD>.md

何时读取 references/

  • 若需要完整房间拓扑、任务卡模板、代谢协议样例:读取 references/architecture.md
  • 若需要发布文案、来源说明、融合声明:读取 references/publishing-note.md
  • 若需要快速命名与路由速查:读取 references/naming-quickref.md
  • 若需要中文发布介绍页:读取 references/release-copy-zh.md
  • 若需要英文发布介绍页:读取 references/release-copy-en.md

交付要求

当使用本技能帮助他人落地时,最终交付应至少包含:

  1. 参考来源说明
  2. 融合逻辑说明
  3. 空间拓扑
  4. 运行层(corridor)
  5. 反思层(reflection wing)
  6. 后台 archivist 方案
  7. 命名与路由规范
  8. 阶段性总控入口

一句话结尾

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。

Usage Guidance
This is a documentation-only skill that outlines folder structure, naming, and workflows for a file-driven multi-agent memory system — it doesn't install software or request secrets. Before using: (1) confirm any agent you give this to is allowed to read/write only the intended project directories (set workspace/file permissions and audit logging); (2) define what counts as 'completed' so an archivist won't process in-progress or sensitive drafts; (3) if you plan autonomous agent runs, restrict which folders and files the agent can access and review runbooks; (4) back up important originals before performing bulk backfills. There are no network endpoints or credential requests embedded in the skill.
Capability Analysis
Type: OpenClaw Skill Name: z1-matrix-memory-palace Version: 1.0.1 The skill bundle provides a conceptual framework and organizational instructions for a file-driven "Memory Palace" memory system for AI agents. It focuses on structuring directories (e.g., grand_hall, chambers, reflection_wing) and managing task cards to reduce token usage and improve long-term knowledge retention. There is no executable code, network activity, or evidence of malicious intent; it is purely a methodology for information architecture within the agent's workspace.
Capability Assessment
Purpose & Capability
The name/description (file-driven Memory Palace + LLM Wiki for multi-agent memory) matches the content: directory topology, naming, workflows and an archivist role. Nothing requested (no env vars, no binaries) is out of scope for a documentation/operational template.
Instruction Scope
SKILL.md instructs creating and operating files/directories, maintaining reflection pages, and running a background archivist that reads completed files. It does not instruct reading unrelated system paths, environment variables, or contacting external endpoints. The only file reads referenced are the packaged references/ files (explicitly enumerated).
Install Mechanism
No install spec and no code files — this is instruction-only. There is nothing that will be downloaded or written by an installer, so installation risk is minimal.
Credentials
The skill declares no required environment variables, credentials, or config paths. All operational guidance concerns local file layouts and naming. There are no requests for unrelated secrets or broad credential access.
Persistence & Privilege
Flags: always:false and user-invocable:true (defaults). The skill does not request persistent system privileges or modifications to other skills. It describes a background 'Archivist' role that reads completed files, which is coherent for the described purpose.
How to Use
  1. Make sure OpenClaw is installed (local or Docker)
  2. Run the install command in chat: /install z1-matrix-memory-palace
  3. After installation, invoke the skill by name or use /z1-matrix-memory-palace
  4. Provide required inputs per the skill's parameter spec and get structured output
Version History
v1.0.1
Add Chinese and English release copy references for publishing and marketing pages.
v1.0.0
Initial release: Memory Palace × LLM Wiki fusion with source attribution, file-driven coordination, reflection wing, archivist, routing and naming spec.
Metadata
Slug z1-matrix-memory-palace
Version 1.0.1
License MIT-0
All-time Installs 0
Active Installs 0
Total Versions 2
Frequently Asked Questions

What is Z1矩阵宫殿记忆法?

Build and operate a file-driven Memory Palace for multi-agent systems, combining a spatial memory shell with a continuously maintained LLM Wiki reflection la... It is an AI Agent Skill for Claude Code / OpenClaw, with 133 downloads so far.

How do I install Z1矩阵宫殿记忆法?

Run "/install z1-matrix-memory-palace" in the OpenClaw or Claude Code chat to install it in one step — no extra setup required.

Is Z1矩阵宫殿记忆法 free?

Yes, Z1矩阵宫殿记忆法 is completely free, licensed under MIT-0. You can download, install and use it at no cost.

Which platforms does Z1矩阵宫殿记忆法 support?

Z1矩阵宫殿记忆法 is cross-platform and runs anywhere OpenClaw / Claude Code is available (cross-platform).

Who created Z1矩阵宫殿记忆法?

It is built and maintained by z1one0415 (@z1one0415); the current version is v1.0.1.

💬 Comments