# claude-memory-kit - Doramagic AI Context Pack

> 定位：安装前体验与判断资产。它帮助宿主 AI 有一个好的开始，但不代表已经安装、执行或验证目标项目。

## 充分原则

- **充分原则，不是压缩原则**：AI Context Pack 应该充分到让宿主 AI 在开工前理解项目价值、能力边界、使用入口、风险和证据来源；它可以分层组织，但不以最短摘要为目标。
- **压缩策略**：只压缩噪声和重复内容，不压缩会影响判断和开工质量的上下文。

## 给宿主 AI 的使用方式

你正在读取 Doramagic 为 claude-memory-kit 编译的 AI Context Pack。请把它当作开工前上下文：帮助用户理解适合谁、能做什么、如何开始、哪些必须安装后验证、风险在哪里。不要声称你已经安装、运行或执行了目标项目。

## Claim 消费规则

- **事实来源**：Repo Evidence + Claim/Evidence Graph；Human Wiki 只提供显著性、术语和叙事结构。
- **事实最低状态**：`supported`
- `supported`：可以作为项目事实使用，但回答中必须引用 claim_id 和证据路径。
- `weak`：只能作为低置信度线索，必须要求用户继续核实。
- `inferred`：只能用于风险提示或待确认问题，不能包装成项目事实。
- `unverified`：不得作为事实使用，应明确说证据不足。
- `contradicted`：必须展示冲突来源，不得替用户强行选择一个版本。

## 它最适合谁

- **正在使用 Claude/Codex/Cursor/Gemini 等宿主 AI 的开发者**：README 或插件配置提到多个宿主 AI。 证据：`README.md` Claim：`clm_0004` supported 0.86
- **希望把专业流程带进宿主 AI 的用户**：仓库包含 Skill 文档。 证据：`plugin/skills/bootstrap/SKILL.md`, `plugin/skills/memory-search/SKILL.md`, `plugin/skills/memory-write/SKILL.md`, `template/.claude/skills/memory-search/SKILL.md` 等 Claim：`clm_0005` supported 0.86

## 它能做什么

- **AI Skill / Agent 指令资产库**（可做安装前预览）：项目包含可被宿主 AI 读取的 Skill 或 Agent 指令文件，可用于把专业流程带入 Claude、Codex、Cursor 等宿主。 证据：`plugin/skills/bootstrap/SKILL.md`, `plugin/skills/memory-search/SKILL.md`, `plugin/skills/memory-write/SKILL.md`, `template/.claude/skills/memory-search/SKILL.md` 等 Claim：`clm_0001` supported 0.86
- **多宿主安装与分发**（需要安装后验证）：项目包含插件或 marketplace 配置，说明它面向一个或多个 AI 宿主的安装和分发。 证据：`.claude-plugin/marketplace.json`, `plugin/.claude-plugin/plugin.json` Claim：`clm_0002` supported 0.86
- **命令行启动或安装流程**（需要安装后验证）：项目文档中存在可执行命令，真实使用需要在本地或宿主环境中运行这些命令。 证据：`README.md` Claim：`clm_0003` supported 0.86

## 怎么开始

- `npm install -g @lh8ppl/claude-memory-kit` 证据：`README.md` Claim：`clm_0006` supported 0.86, `clm_0009` supported 0.86
- `/plugin marketplace add LH8PPL/claude-memory-kit   # add this repo as a plugin source (once per machine)` 证据：`README.md` Claim：`clm_0007` supported 0.86
- `/plugin install claude-memory-kit                  # install the global machinery — hooks + skills (once per machine)` 证据：`README.md` Claim：`clm_0008` supported 0.86
- `npm install -g @lh8ppl/claude-memory-kit@latest   # 1. update the global cmk CLI` 证据：`README.md` Claim：`clm_0009` supported 0.86

## 继续前判断卡

- **当前建议**：仅建议沙盒试装
- **为什么**：项目存在安装命令、宿主配置或本地写入线索，不建议直接进入主力环境，应先在隔离环境试装。

### 30 秒判断

- **现在怎么做**：仅建议沙盒试装
- **最小安全下一步**：先跑 Prompt Preview；若仍要安装，只在隔离环境试装
- **先别相信**：真实输出质量不能在安装前相信。
- **继续会触碰**：命令执行、宿主 AI 配置、本地环境或项目文件

### 现在可以相信

- **适合人群线索：正在使用 Claude/Codex/Cursor/Gemini 等宿主 AI 的开发者**（supported）：有 supported claim 或项目证据支撑，但仍不等于真实安装效果。 证据：`README.md` Claim：`clm_0004` supported 0.86
- **适合人群线索：希望把专业流程带进宿主 AI 的用户**（supported）：有 supported claim 或项目证据支撑，但仍不等于真实安装效果。 证据：`plugin/skills/bootstrap/SKILL.md`, `plugin/skills/memory-search/SKILL.md`, `plugin/skills/memory-write/SKILL.md`, `template/.claude/skills/memory-search/SKILL.md` 等 Claim：`clm_0005` supported 0.86
- **能力存在：AI Skill / Agent 指令资产库**（supported）：可以相信项目包含这类能力线索；是否适合你的具体任务仍要试用或安装后验证。 证据：`plugin/skills/bootstrap/SKILL.md`, `plugin/skills/memory-search/SKILL.md`, `plugin/skills/memory-write/SKILL.md`, `template/.claude/skills/memory-search/SKILL.md` 等 Claim：`clm_0001` supported 0.86
- **能力存在：多宿主安装与分发**（supported）：可以相信项目包含这类能力线索；是否适合你的具体任务仍要试用或安装后验证。 证据：`.claude-plugin/marketplace.json`, `plugin/.claude-plugin/plugin.json` Claim：`clm_0002` supported 0.86
- **能力存在：命令行启动或安装流程**（supported）：可以相信项目包含这类能力线索；是否适合你的具体任务仍要试用或安装后验证。 证据：`README.md` Claim：`clm_0003` supported 0.86
- **存在 Quick Start / 安装命令线索**（supported）：可以相信项目文档出现过启动或安装入口；不要因此直接在主力环境运行。 证据：`README.md` Claim：`clm_0006` supported 0.86, `clm_0009` supported 0.86

### 现在还不能相信

- **真实输出质量不能在安装前相信。**（unverified）：Prompt Preview 只能展示引导方式，不能证明真实项目中的结果质量。
- **宿主 AI 版本兼容性不能在安装前相信。**（unverified）：Claude、Cursor、Codex、Gemini 等宿主加载规则和版本差异必须在真实环境验证。
- **不会污染现有宿主 AI 行为，不能直接相信。**（inferred）：Skill、plugin、AGENTS/CLAUDE/GEMINI 指令可能改变宿主 AI 的默认行为。 证据：`.claude-plugin/marketplace.json`, `CLAUDE.md`, `docs/research/2026-05-21-claude-mem-architecture.md`, `plugin/.claude-plugin/plugin.json` 等
- **可安全回滚不能默认相信。**（unverified）：除非项目明确提供卸载和恢复说明，否则必须先在隔离环境验证。
- **真实安装后是否与用户当前宿主 AI 版本兼容？**（unverified）：兼容性只能通过实际宿主环境验证。 证据：`.claude-plugin/marketplace.json`, `plugin/.claude-plugin/plugin.json`
- **项目输出质量是否满足用户具体任务？**（unverified）：安装前预览只能展示流程和边界，不能替代真实评测。
- **安装命令是否需要网络、权限或全局写入？**（unverified）：这影响企业环境和个人环境的安装风险。 证据：`README.md`

### 继续会触碰什么

- **命令执行**：包管理器、网络下载、本地插件目录、项目配置或用户主目录。 原因：运行第一条命令就可能产生环境改动；必须先判断是否值得跑。 证据：`README.md`
- **宿主 AI 配置**：Claude/Codex/Cursor/Gemini/OpenCode 等宿主的 plugin、Skill 或规则加载配置。 原因：宿主配置会改变 AI 后续工作方式，可能和用户已有规则冲突。 证据：`.claude-plugin/marketplace.json`, `CLAUDE.md`, `docs/research/2026-05-21-claude-mem-architecture.md`, `plugin/.claude-plugin/plugin.json` 等
- **本地环境或项目文件**：安装结果、插件缓存、项目配置或本地依赖目录。 原因：安装前无法证明写入范围和回滚方式，需要隔离验证。 证据：`.claude-plugin/marketplace.json`, `README.md`, `plugin/.claude-plugin/plugin.json`
- **宿主 AI 上下文**：AI Context Pack、Prompt Preview、Skill 路由、风险规则和项目事实。 原因：导入上下文会影响宿主 AI 后续判断，必须避免把未验证项包装成事实。

### 最小安全下一步

- **先跑 Prompt Preview**：用安装前交互式试用判断工作方式是否匹配，不需要授权或改环境。（适用：任何项目都适用，尤其是输出质量未知时。）
- **只在隔离目录或测试账号试装**：避免安装命令污染主力宿主 AI、真实项目或用户主目录。（适用：存在命令执行、插件配置或本地写入线索时。）
- **先备份宿主 AI 配置**：Skill、plugin、规则文件可能改变 Claude/Cursor/Codex 的默认行为。（适用：存在插件 manifest、Skill 或宿主规则入口时。）
- **安装后只验证一个最小任务**：先验证加载、兼容、输出质量和回滚，再决定是否深用。（适用：准备从试用进入真实工作流时。）

### 退出方式

- **保留安装前状态**：记录原始宿主配置和项目状态，后续才能判断是否可恢复。
- **准备移除宿主 plugin / Skill / 规则入口**：如果试装后行为异常，可以把宿主 AI 恢复到试装前状态。
- **记录安装命令和写入路径**：没有明确卸载说明时，至少要知道哪些目录或配置需要手动清理。
- **如果没有回滚路径，不进入主力环境**：不可回滚是继续前阻断项，不应靠信任或运气继续。

## 哪些只能预览

- 解释项目适合谁和能做什么
- 基于项目文档演示典型对话流程
- 帮助用户判断是否值得安装或继续研究

## 哪些必须安装后验证

- 真实安装 Skill、插件或 CLI
- 执行脚本、修改本地文件或访问外部服务
- 验证真实输出质量、性能和兼容性

## 边界与风险判断卡

- **把安装前预览误认为真实运行**：用户可能高估项目已经完成的配置、权限和兼容性验证。 处理方式：明确区分 prompt_preview_can_do 与 runtime_required。 Claim：`clm_0010` inferred 0.45
- **宿主 AI 插件或 Skill 规则冲突**：新规则可能改变用户现有宿主 AI 的工作方式。 处理方式：安装前先检查插件 manifest 和 Skill 文件，必要时隔离测试。 证据：`.claude-plugin/marketplace.json`, `plugin/.claude-plugin/plugin.json` Claim：`clm_0011` supported 0.86
- **命令执行会修改本地环境**：安装命令可能写入用户主目录、宿主插件目录或项目配置。 处理方式：先在隔离环境或测试账号中运行。 证据：`README.md` Claim：`clm_0012` supported 0.86
- **源文档冲突：skill_count**：项目文档中存在数量表述不一致，AI Context Pack 必须提示用户不要把单一数字当作已验证事实。 处理方式：在 Human Manual 和 AI Context Pack 中共同标记为待核实，而不是强行选择一个数字。 证据：`docs/research/2026-05-21-claude-mem-architecture.md`, `docs/research/2026-06-01-how-products-implement-skills.md`, `context/memory/project_1-gate-run-verification-expected-health-checks-and-session.md` Claim：`clm_0013` supported 0.86, `clm_0017` contradicted 0.20
- **源文件冲突 skill_count**：发现多个值 `15, 195`，应在真实使用前核实。
- **待确认**：真实安装后是否与用户当前宿主 AI 版本兼容？。原因：兼容性只能通过实际宿主环境验证。
- **待确认**：项目输出质量是否满足用户具体任务？。原因：安装前预览只能展示流程和边界，不能替代真实评测。
- **待确认**：安装命令是否需要网络、权限或全局写入？。原因：这影响企业环境和个人环境的安装风险。

## 开工前工作上下文

### 加载顺序

- 先读取 how_to_use.host_ai_instruction，建立安装前判断资产的边界。
- 读取 claim_graph_summary，确认事实来自 Claim/Evidence Graph，而不是 Human Wiki 叙事。
- 再读取 intended_users、capabilities 和 quick_start_candidates，判断用户是否匹配。
- 需要执行具体任务时，优先查 role_skill_index，再查 evidence_index。
- 遇到真实安装、文件修改、网络访问、性能或兼容性问题时，转入 risk_card 和 boundaries.runtime_required。

### 任务路由

- **AI Skill / Agent 指令资产库**：先基于 role_skill_index / evidence_index 帮用户挑选可用角色、Skill 或工作流。 边界：可做安装前 Prompt 体验。 证据：`plugin/skills/bootstrap/SKILL.md`, `plugin/skills/memory-search/SKILL.md`, `plugin/skills/memory-write/SKILL.md`, `template/.claude/skills/memory-search/SKILL.md` 等 Claim：`clm_0001` supported 0.86
- **多宿主安装与分发**：先说明这是安装后验证能力，再给出安装前检查清单。 边界：必须真实安装或运行后验证。 证据：`.claude-plugin/marketplace.json`, `plugin/.claude-plugin/plugin.json` Claim：`clm_0002` supported 0.86
- **命令行启动或安装流程**：先说明这是安装后验证能力，再给出安装前检查清单。 边界：必须真实安装或运行后验证。 证据：`README.md` Claim：`clm_0003` supported 0.86

### 上下文规模

- 文件总数：1156
- 重要文件覆盖：40/1156
- 证据索引条目：80
- 角色 / Skill 条目：5

### 证据不足时的处理

- **missing_evidence**：说明证据不足，要求用户提供目标文件、README 段落或安装后验证记录；不要补全事实。
- **out_of_scope_request**：说明该任务超出当前 AI Context Pack 证据范围，并建议用户先查看 Human Manual 或真实安装后验证。
- **runtime_request**：给出安装前检查清单和命令来源，但不要替用户执行命令或声称已执行。
- **source_conflict**：同时展示冲突来源，标记为待核实，不要强行选择一个版本。

## Prompt Recipes

### 适配判断

- 目标：判断这个项目是否适合用户当前任务。
- 预期输出：适配结论、关键理由、证据引用、安装前可预览内容、必须安装后验证内容、下一步建议。

```text
请基于 claude-memory-kit 的 AI Context Pack，先问我 3 个必要问题，然后判断它是否适合我的任务。回答必须包含：适合谁、能做什么、不能做什么、是否值得安装、证据来自哪里。所有项目事实必须引用 evidence_refs、source_paths 或 claim_id。
```

### 安装前体验

- 目标：让用户在安装前感受核心工作流，同时避免把预览包装成真实能力或营销承诺。
- 预期输出：一段带边界标签的体验剧本、安装后验证清单和谨慎建议；不含真实运行承诺或强营销表述。

```text
请把 claude-memory-kit 当作安装前体验资产，而不是已安装工具或真实运行环境。

请严格输出四段：
1. 先问我 3 个必要问题。
2. 给出一段“体验剧本”：用 [安装前可预览]、[必须安装后验证]、[证据不足] 三种标签展示它可能如何引导工作流。
3. 给出安装后验证清单：列出哪些能力只有真实安装、真实宿主加载、真实项目运行后才能确认。
4. 给出谨慎建议：只能说“值得继续研究/试装”“先补充信息后再判断”或“不建议继续”，不得替项目背书。

硬性边界：
- 不要声称已经安装、运行、执行测试、修改文件或产生真实结果。
- 不要写“自动适配”“确保通过”“完美适配”“强烈建议安装”等承诺性表达。
- 如果描述安装后的工作方式，必须使用“如果安装成功且宿主正确加载 Skill，它可能会……”这种条件句。
- 体验剧本只能写成“示例台词/假设流程”：使用“可能会询问/可能会建议/可能会展示”，不要写“已写入、已生成、已通过、正在运行、正在生成”。
- Prompt Preview 不负责给安装命令；如用户准备试装，只能提示先阅读 Quick Start 和 Risk Card，并在隔离环境验证。
- 所有项目事实必须来自 supported claim、evidence_refs 或 source_paths；inferred/unverified 只能作风险或待确认项。

```

### 角色 / Skill 选择

- 目标：从项目里的角色或 Skill 中挑选最匹配的资产。
- 预期输出：候选角色或 Skill 列表，每项包含适用场景、证据路径、风险边界和是否需要安装后验证。

```text
请读取 role_skill_index，根据我的目标任务推荐 3-5 个最相关的角色或 Skill。每个推荐都要说明适用场景、可能输出、风险边界和 evidence_refs。
```

### 风险预检

- 目标：安装或引入前识别环境、权限、规则冲突和质量风险。
- 预期输出：环境、权限、依赖、许可、宿主冲突、质量风险和未知项的检查清单。

```text
请基于 risk_card、boundaries 和 quick_start_candidates，给我一份安装前风险预检清单。不要替我执行命令，只说明我应该检查什么、为什么检查、失败会有什么影响。
```

### 宿主 AI 开工指令

- 目标：把项目上下文转成一次对话开始前的宿主 AI 指令。
- 预期输出：一段边界明确、证据引用明确、适合复制给宿主 AI 的开工前指令。

```text
请基于 claude-memory-kit 的 AI Context Pack，生成一段我可以粘贴给宿主 AI 的开工前指令。这段指令必须遵守 not_runtime=true，不能声称项目已经安装、运行或产生真实结果。
```

## 角色 / Skill 索引

- 共索引 5 个角色 / Skill / 项目文档条目。

- **bootstrap**（skill）： 激活提示：当用户任务与“bootstrap”描述的流程高度相关时，先用它做安装前体验，再决定是否安装。 证据：`plugin/skills/bootstrap/SKILL.md`
- **memory-search**（skill）：- 激活提示：当用户任务与“memory-search”描述的流程高度相关时，先用它做安装前体验，再决定是否安装。 证据：`plugin/skills/memory-search/SKILL.md`
- **memory-write**（skill）：- 激活提示：当用户任务与“memory-write”描述的流程高度相关时，先用它做安装前体验，再决定是否安装。 证据：`plugin/skills/memory-write/SKILL.md`
- **memory-search**（skill）：- 激活提示：当用户任务与“memory-search”描述的流程高度相关时，先用它做安装前体验，再决定是否安装。 证据：`template/.claude/skills/memory-search/SKILL.md`
- **memory-write**（skill）：- 激活提示：当用户任务与“memory-write”描述的流程高度相关时，先用它做安装前体验，再决定是否安装。 证据：`template/.claude/skills/memory-write/SKILL.md`

## 证据索引

- 共索引 80 条证据。

- **claude-memory-kit — documentation system**（documentation）：claude-memory-kit — documentation system 证据：`docs/README.md`
- **Conversation log**（documentation）：⛔ RETIRED 2026-05-30. No longer maintained. It went dormant after 2026-05-22 the whole v0.1.x → v0.2 stretch was never logged here , and its role is now covered by two live surfaces: - docs/journey/DECISION-LOG.md ../journey/DECISION-LOG.md — what was decided + why append-only, per CLAUDE.md "Decision-log discipline" . - docs/journey/build-log.md ../journey/build-log.md — narrative / messy path / per-PR retrospectives . The two 2026-05-2x.md files are kept as the historical trail. Do not add new entries here — use the two surfaces above and keep THOSE current. The lesson of this retirement: a doc surface left to go dormant becomes a liability. If you ever reach for a third narrative surface… 证据：`archive/docs/conversation-log/README.md`
- **Architectural Decision Records ADRs**（documentation）：Architectural Decision Records ADRs 证据：`docs/adr/README.md`
- **Process documentation**（documentation）：How we work on claude-memory-kit . Each process doc is a self-contained markdown file explaining one part of our methodology — the spec workflow, research conventions, git discipline, etc. 证据：`docs/process/README.md`
- **Sources — deep-dive notes**（documentation）：Detailed notes on individual external sources we've consulted. Each source gets its own markdown file with the substance distilled — so a future reader or future Claude session, or personal-wiki ingest gets the value of the source without re-reading the original. 证据：`docs/sources/README.md`
- **How it feels**（documentation）：Persistent, per-project memory for Claude Code — plain markdown, committed with your code, recalled by meaning. 证据：`README.md`
- **claude-memory-kit Claude Code plugin**（documentation）：claude-memory-kit Claude Code plugin 证据：`plugin/README.md`
- **cmk-canonicalize Python**（documentation）：Python parallel to the Node @lh8ppl/cmk-canonicalize package. Same public API, byte-identical output against ../fixtures/canonicalize-vectors.json ../fixtures/canonicalize-vectors.json . 证据：`python/README.md`
- **@lh8ppl/claude-memory-kit**（documentation）：cmk — the CLI for claude-memory-kit https://github.com/LH8PPL/claude-memory-kit , a per-project, in-repo memory system for Claude Code https://docs.claude.com/en/docs/claude-code . It fixes Claude's per-session amnesia so you don't have to re-tell the backstory every time you start a new session. 证据：`packages/cli/README.md`
- **Package**（package_manifest）：{ "name": "@claude-memory-kit/repo", "version": "0.0.1", "private": true, "description": "Per-project, in-repo memory system for Claude Code. Monorepo root for development; published packages live under packages/.", "type": "module", "workspaces": "packages/ " , "scripts": { "lint": "node scripts/validate-test-ids.mjs && node scripts/validate-template.mjs && node scripts/validate-exit-doors.mjs && node scripts/validate-references.mjs && node scripts/validate-spawn-discipline.mjs && node scripts/validate-numbering-gaps.mjs && node scripts/validate-composition.mjs && node scripts/validate-budget-pairs.mjs && node scripts/validate-prompt-assertions.mjs && node scripts/validate-skill-allowlist.… 证据：`package.json`
- **Package**（package_manifest）：{ "name": "@lh8ppl/cmk-canonicalize", "version": "0.1.0", "description": "Deterministic text canonicalization + content-addressed ID generation for claude-memory-kit. Node + Python parity package; output is byte-identical across both implementations.", "type": "module", "exports": { ".": "./src/index.mjs" }, "files": "src/", "README.md" , "scripts": { "test": "vitest run" }, "engines": { "node": " =20" }, "license": "MIT", "repository": { "type": "git", "url": "https://github.com/LH8PPL/claude-memory-kit.git", "directory": "packages/canonicalize" }, "bugs": { "url": "https://github.com/LH8PPL/claude-memory-kit/issues" }, "keywords": "claude", "claude-code", "memory", "canonicalize", "conten… 证据：`packages/canonicalize/package.json`
- **Package**（package_manifest）：{ "name": "@lh8ppl/claude-memory-kit", "version": "0.4.0", "description": "cmk — the CLI for claude-memory-kit. Per-project, in-repo memory system for Claude Code.", "type": "module", "bin": { "cmk": "./bin/cmk.mjs", "cmk-daily-distill": "./bin/cmk-daily-distill.mjs", "cmk-weekly-curate": "./bin/cmk-weekly-curate.mjs", "cmk-compress-lazy": "./bin/cmk-compress-lazy.mjs", "cmk-inject-context": "./bin/cmk-inject-context.mjs", "cmk-capture-prompt": "./bin/cmk-capture-prompt.mjs", "cmk-observe-edit": "./bin/cmk-observe-edit.mjs", "cmk-capture-turn": "./bin/cmk-capture-turn.mjs", "cmk-compress-session": "./bin/cmk-compress-session.mjs", "cmk-guard-memory": "./bin/cmk-guard-memory.mjs" }, "files":… 证据：`packages/cli/package.json`
- **claude-memory-kit — project CLAUDE.md**（documentation）：claude-memory-kit — project CLAUDE.md 证据：`CLAUDE.md`
- **cmk CLI reference**（documentation）：Complete reference for the cmk command installed by npm install -g @lh8ppl/claude-memory-kit . Run cmk --help or cmk --help for the live version. 证据：`docs/CLI.md`
- **cmk — full test + cut gate KIRO-CLI / terminal**（documentation）：cmk — full test + cut gate KIRO-CLI / terminal 证据：`docs/process/cut-gate-kiro-cli.md`
- **Contributing to claude-memory-kit**（documentation）：Thanks for your interest in improving the kit itself as opposed to installing it in your own project . This guide covers the development workflow, testing, and the conventions the codebase follows. 证据：`CONTRIBUTING.md`
- **bootstrap**（skill_instruction）：Scaffold the memory-system files into the user's current project. The plugin itself ships with the hooks and skill, but the user's project needs its own context/ directory tree, scripts/ for the auto-extract helper to read project-relative paths, and the cron job declarations. 证据：`plugin/skills/bootstrap/SKILL.md`
- **Recalling from deep memory**（skill_instruction）：You are a memory-retrieval agent. Search the kit's memory archive for: $ARGUMENTS 证据：`plugin/skills/memory-search/SKILL.md`
- **Capturing durable memory**（skill_instruction）：Durable facts — preferences, decisions, environment state — are saved through the kit's safe write path Poison Guard secret screening + home-path sanitization + dedup + conflict detection . 证据：`plugin/skills/memory-write/SKILL.md`
- **Recalling from deep memory**（skill_instruction）：You are a memory-retrieval agent. Search the kit's memory archive for: $ARGUMENTS 证据：`template/.claude/skills/memory-search/SKILL.md`
- **Capturing durable memory**（skill_instruction）：Durable facts — preferences, decisions, environment state — are saved through the kit's safe write path Poison Guard secret screening + home-path sanitization + dedup + conflict detection . 证据：`template/.claude/skills/memory-write/SKILL.md`
- **Marketplace**（structured_config）：{ "name": "claude-memory-kit", "owner": { "name": "Lior Hollander" }, "description": "Per-project, in-repo memory system for Claude Code — the cross-session memory the kit's whole thesis is about.", "plugins": { "name": "claude-memory-kit", "source": "./plugin", "description": "Frozen-snapshot session loads, auto-extract Stop hook, memory-write + bootstrap skills, optional memsearch semantic recall, optional auto-curation crons.", "version": "0.1.1", "author": { "name": "Lior Hollander" }, "license": "MIT" } } 证据：`.claude-plugin/marketplace.json`
- **Plugin**（structured_config）：{ "name": "claude-memory-kit", "description": "Per-project in-repo memory system for Claude Code. Frozen-snapshot session loads, auto-extract Stop hook, memory-write skill, keyword search, optional auto-curation crons.", "version": "0.1.1", "author": { "name": "Lior Hollander" }, "homepage": "https://github.com/LH8PPL/claude-memory-kit", "repository": "https://github.com/LH8PPL/claude-memory-kit", "license": "MIT" } 证据：`plugin/.claude-plugin/plugin.json`
- **License**（source_file）：Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files the "Software" , to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions: 证据：`LICENSE`
- **Changelog**（documentation）：All notable changes to claude-memory-kit are documented in this file. 证据：`CHANGELOG.md`
- **Health checks**（documentation）：Eight yes/no checks cmk doctor runs against the kit installation. Each has a self-repair path. The two memsearch checks — formerly HC-1 "installed" + HC-7 "reachable" — were removed in Task 120: the kit ships keyword-only and the Layer-5b semantic backend is not yet shipped; the backend choice is deferred per design §9.3.1. 证据：`HEALTH-CHECKS.md`
- **Bootstrap prompt for a new Claude Code session**（documentation）：Bootstrap prompt for a new Claude Code session 证据：`docs/BOOTSTRAP.md`
- **DOCUMENTATION-MAP — the three files that matter, and everything that orbits them**（documentation）：DOCUMENTATION-MAP — the three files that matter, and everything that orbits them 证据：`docs/DOCUMENTATION-MAP.md`
- **MCP tools mcp cmk**（documentation）：The kit ships an MCP server so Claude can run every memory operation in conversation — you never type cmk . cmk install registers it writes .mcp.json and allow-lists mcp cmk in .claude/settings.json , so the tools are available prompt-free the moment you open Claude Code. The server is stdio-only and local — no network listener, no auth surface. 证据：`docs/MCP.md`
- **Release plan — task → version map**（documentation）：Single source of truth for WHICH version each task ships in. When deciding "what version does task X go in?", this file answers it. tasks.md holds task DETAIL + checkbox state; this file holds the version LANE. Update it in the same batch whenever a task's target version changes. 证据：`docs/RELEASE-PLAN.md`
- **Master sources index**（documentation）：Every URL, paper, repo, video, and blog post cited anywhere in claude-memory-kit . The authoritative reference list. When you need a citation, link here. 证据：`docs/SOURCES.md`
- **Session log — 2026-05-21**（documentation）：Today's session covered the entire bootstrap of claude-memory-kit from a continuation of work on youtube-to-slide . Major outcomes: kit v0.0.1 shipped to private GitHub; claude-mem discovered and evaluated; Kiro-style spec workflow adopted; requirements.md drafted; 7 open questions resolved; both Deep Research prompts Option A and Option B dispatched. 证据：`archive/docs/conversation-log/2026-05-21.md`
- **Session log — 2026-05-22**（documentation）：Today's session: Option-B Deep Research returned from Claude.ai; we analyzed the findings, identified bank/airgap injection as a Claude.ai memory artifact, and established the full documentation discipline this conversation-log entry is one product of that decision . 证据：`archive/docs/conversation-log/2026-05-22.md`
- **Phase 3 plan — folded into the spine pointer only**（documentation）：Phase 3 plan — folded into the spine pointer only 证据：`archive/docs/journey/PHASE-3-PLAN.md`
- **Feedback Use Mcp Mk Remember Not Bash Cli**（documentation）：When capturing memory during a session on THIS repo which dogfoods the kit , use the mk remember MCP tool — NOT a Bash shell-out to node packages/cli/bin/cmk.mjs remember or cmk remember . The MCP tool is the in-conversation surface the kit is built to exercise; reaching for the raw CLI bypasses it and is the make-the-user-the-orchestrator anti-pattern. 证据：`context/memory/feedback_use-mcp-mk-remember-not-bash-cli.md`
- **Project The Fix Use Cmk Remember Cli Not Mk Remember Mcp In Kiro Cli**（documentation）：★★★ THE FIX the user's insight, 2026-06-24 : use the CLI COMMAND cmk remember instead of the MCP TOOL mk remember in kiro-cli — it BYPASSES the Kiro MCP bug 5873 entirely. PROVEN: cmk remember "..." in gate4 → "saved to P/MEMORY.md P-7DXHJMW9 ", audit 1→2, LANDS correctly. The Kiro bug 5873 only affects MCP TOOLS kiro doesn't wire MCP tools to custom agents' LLM ; cmk remember is a plain SHELL COMMAND — kiro runs shell commands fine the hooks prove it: cmd.exe /c cmk hook works . So the explicit-capture path works via the CLI, just not via MCP. BONUS: the scaffolded memory-write SKILL.md ALREADY has both surfaces — "Preferred: the cmk MCP tools when connected " + "Fallback: the cmk CLI when… 证据：`context/memory/project_the-fix-use-cmk-remember-cli-not-mk-remember-mcp-in-kiro-cli.md`
- **ADR-0001 — Build claude-memory-kit as a separate project, not by forking youtube-to-slide**（documentation）：ADR-0001 — Build claude-memory-kit as a separate project, not by forking youtube-to-slide 证据：`docs/adr/0001-separate-project-not-fork-youtube-to-slide.md`
- **ADR-0002 — Markdown files as the source of truth; SQLite/vector DBs are regenerable caches**（documentation）：ADR-0002 — Markdown files as the source of truth; SQLite/vector DBs are regenerable caches 证据：`docs/adr/0002-markdown-source-of-truth-over-opaque-db.md`
- **ADR-0003 — Per-project memory in v0.1; three-tier scope user / project / local lands in v0.1.0 design**（documentation）：ADR-0003 — Per-project memory in v0.1; three-tier scope user / project / local lands in v0.1.0 design 证据：`docs/adr/0003-per-project-with-future-cross-project-tier.md`
- **ADR-0004 — Spec-driven development using a Kiro-adapted workflow requirements → design → tasks**（documentation）：ADR-0004 — Spec-driven development using a Kiro-adapted workflow requirements → design → tasks 证据：`docs/adr/0004-spec-driven-development-kiro-style.md`
- **ADR-0005 — Ship three install paths — bash script, PowerShell script, Claude Code plugin**（documentation）：ADR-0005 — Ship three install paths — bash script, PowerShell script, Claude Code plugin 证据：`docs/adr/0005-three-install-paths.md`
- **ADR-0006 — Lifecycle hook architecture — initial 6 hooks, revised to 5+1 after Option-B research**（documentation）：ADR-0006 — Lifecycle hook architecture — initial 6 hooks, revised to 5+1 after Option-B research 证据：`docs/adr/0006-lifecycle-hooks-architecture.md`
- **ADR-0007 — Content-addressed citation IDs — 8-char base32 SHA-256 with tier prefix**（documentation）：ADR-0007 — Content-addressed citation IDs — 8-char base32 SHA-256 with tier prefix 证据：`docs/adr/0007-content-addressed-citation-ids.md`
- **ADR-0008 — Bank / air-gapped deployment deferred to v0.2+ but compressor designed pluggably**（documentation）：ADR-0008 — Bank / air-gapped deployment deferred to v0.2+ but compressor designed pluggably 证据：`docs/adr/0008-bank-airgap-deferred-to-future-version.md`
- **ADR-0009 — Provenance frontmatter on every observation**（documentation）：ADR-0009 — Provenance frontmatter on every observation 证据：`docs/adr/0009-provenance-frontmatter-per-observation.md`
- **ADR-0010 — Raw transcripts preserved indefinitely**（documentation）：ADR-0010 — Raw transcripts preserved indefinitely 证据：`docs/adr/0010-raw-transcripts-preserved-indefinitely.md`
- **ADR-0011 — Coexistence strategy with Anthropic's built-in Auto Memory Claude Code v2.1.59+**（documentation）：ADR-0011 — Coexistence strategy with Anthropic's built-in Auto Memory Claude Code v2.1.59+ 证据：`docs/adr/0011-coexistence-with-anthropic-auto-memory.md`
- **ADR-0012 — Publish v0.1.0 under @lh8ppl scope; defer the cross-agent product name to v0.2**（documentation）：ADR-0012 — Publish v0.1.0 under @lh8ppl scope; defer the cross-agent product name to v0.2 证据：`docs/adr/0012-npm-publish-name-and-cross-agent-future.md`
- **0013 — Package security posture and CI provenance publish**（documentation）：0013 — Package security posture and CI provenance publish 证据：`docs/adr/0013-package-security-posture-and-ci-provenance-publish.md`
- **ADR-0014 — Unify CLI + MCP over one in-process memory-op core**（documentation）：ADR-0014 — Unify CLI + MCP over one in-process memory-op core 证据：`docs/adr/0014-unify-cli-mcp-shared-core.md`
- **ADR-0015 — Layer-5b semantic backend: sqlite-vec inside the existing index + an optional local ONNX embedder bge-base-e…**（documentation）：ADR-0015 — Layer-5b semantic backend: sqlite-vec inside the existing index + an optional local ONNX embedder bge-base-en-v1.5 证据：`docs/adr/0015-semantic-backend-sqlite-vec-plus-local-onnx-embedder.md`
- **Live-test findings — scenarios 3-7 completes the original 7-scenario plan**（documentation）：Live-test findings — scenarios 3-7 completes the original 7-scenario plan 证据：`docs/journey/2026-05-26-live-test-findings-scenarios-3-7.md`
- **Live-test findings — plugin-layout path mismatch**（documentation）：Live-test findings — plugin-layout path mismatch 证据：`docs/journey/2026-05-26-live-test-findings.md`
- **Snapshot cap coordination — completing the user-tier-cap structural fix**（documentation）：Snapshot cap coordination — completing the user-tier-cap structural fix 证据：`docs/journey/2026-05-26-snapshot-cap-coordination.md`
- **User-tier cap-truncation fix — per-tier byte budgets**（documentation）：User-tier cap-truncation fix — per-tier byte budgets 证据：`docs/journey/2026-05-26-user-tier-cap-fix.md`
- **Decision log — running paper trail**（documentation）：Purpose. An append-only, chronological record of every decision / issue / bug / fix / pivot so that a future session after auto-compact or a premature end can reconstruct what was decided and why without re-deriving it from faded context. This file exists because on 2026-05-30 the assistant recommended re-opening an already-settled decision build v0.2 vs. a v0.1.3 patch-sweep — reasoning from in-context memory instead of the durable record. The user: "we will not have this problem again that you decide something against an already decision line we are already into." Binding usage see CLAUDE.md "Decision-log discipline" : 1. Before re-opening ANY decision — search this file + v0.1.1-self-tes… 证据：`docs/journey/DECISION-LOG.md`
- **Resume here**（documentation）：This file is a pointer, not state. Where we are + what's next lives in the Spine — read it there, not here. Per ../DOCUMENTATION-MAP.md ../DOCUMENTATION-MAP.md , current project state lives ONLY in requirements/design/tasks; a handoff file that carries its own state always drifts this one did — 425 lines of stacked, superseded blocks, now folded into tasks.md . 证据：`docs/journey/RESUME-HERE-2026-05-28.md`
- **Building claude-memory-kit v0.1.0 — journey log**（documentation）：Building claude-memory-kit v0.1.0 — journey log 证据：`docs/journey/build-log.md`
- **live-test findings — 2026-06-04**（documentation）：27 pass · 2 fail · 0 skip · claude 2.1.160 Claude Code 证据：`docs/journey/live-test-runs/live-test-2026-06-04T090516Z.md`
- **live-test findings — 2026-06-11**（documentation）：28 pass · 1 fail · 0 skip gate-fail 1; 0 known-variable recall probe fail — Task 75 · claude 2.1.170 Claude Code 证据：`docs/journey/live-test-runs/live-test-2026-06-11T114519Z.md`
- 其余 20 条证据见 `AI_CONTEXT_PACK.json` 或 `EVIDENCE_INDEX.json`。

## 宿主 AI 必须遵守的规则

- **把本资产当作开工前上下文，而不是运行环境。**：AI Context Pack 只包含证据化项目理解，不包含目标项目的可执行状态。 证据：`docs/README.md`, `archive/docs/conversation-log/README.md`, `docs/adr/README.md`
- **回答用户时区分可预览内容与必须安装后才能验证的内容。**：安装前体验的消费者价值来自降低误装和误判，而不是伪装成真实运行。 证据：`docs/README.md`, `archive/docs/conversation-log/README.md`, `docs/adr/README.md`

## 用户开工前应该回答的问题

- 你准备在哪个宿主 AI 或本地环境中使用它？
- 你只是想先体验工作流，还是准备真实安装？
- 你最在意的是安装成本、输出质量、还是和现有规则的冲突？

## 验收标准

- 所有能力声明都能回指到 evidence_refs 中的文件路径。
- AI_CONTEXT_PACK.md 没有把预览包装成真实运行。
- 用户能在 3 分钟内看懂适合谁、能做什么、如何开始和风险边界。

---

## Doramagic Context Augmentation

下面内容用于强化 Repomix/AI Context Pack 主体。Human Manual 只提供阅读骨架；踩坑日志会被转成宿主 AI 必须遵守的工作约束。

## Human Manual 骨架

使用规则：这里只是项目阅读路线和显著性信号，不是事实权威。具体事实仍必须回到 repo evidence / Claim Graph。

宿主 AI 硬性规则：
- 不得把页标题、章节顺序、摘要或 importance 当作项目事实证据。
- 解释 Human Manual 骨架时，必须明确说它只是阅读路线/显著性信号。
- 能力、安装、兼容性、运行状态和风险判断必须引用 repo evidence、source path 或 Claim Graph。

- **Overview and Three-Tier Memory Model**：importance `high`
  - source_paths: README.md, ARCHITECTURE.md, docs/README.md, CLAUDE.md, context/MEMORY.md
- **Installation, CLI Reference, and Agent Surfaces (Claude Code + Kiro)**：importance `high`
  - source_paths: README.md, QUICKSTART.md, docs/CLI.md, docs/MCP.md, docs/BOOTSTRAP.md
- **Architecture, Components, and Data Flow**：importance `high`
  - source_paths: ARCHITECTURE.md, docs/adr/0001-separate-project-not-fork-youtube-to-slide.md, docs/adr/0006-lifecycle-hooks-architecture.md, docs/adr/0009-provenance-frontmatter-per-observation.md, docs/adr/0014-unify-cli-mcp-shared-core.md
- **Operations, Health Checks, Compression, and Common Failure Modes**：importance `high`
  - source_paths: HEALTH-CHECKS.md, docs/RELEASE-PLAN.md, docs/process/cut-gate.md, docs/process/cut-gate-kiro.md, docs/process/cut-gate-kiro-cli.md

## Repo Inspection Evidence / 源码检查证据

- repo_clone_verified: true
- repo_inspection_verified: true
- repo_commit: `f47adec42766fbc09a3dd8b5136fcf0530258310`
- inspected_files: `README.md`, `package.json`, `docs/BOOTSTRAP.md`, `docs/CLI.md`, `docs/DOCUMENTATION-MAP.md`, `docs/MCP.md`, `docs/README.md`, `docs/RELEASE-PLAN.md`, `docs/SOURCES.md`, `docs/adr/0001-separate-project-not-fork-youtube-to-slide.md`, `docs/adr/0002-markdown-source-of-truth-over-opaque-db.md`, `docs/adr/0003-per-project-with-future-cross-project-tier.md`, `docs/adr/0004-spec-driven-development-kiro-style.md`, `docs/adr/0005-three-install-paths.md`, `docs/adr/0006-lifecycle-hooks-architecture.md`, `docs/adr/0007-content-addressed-citation-ids.md`, `docs/adr/0008-bank-airgap-deferred-to-future-version.md`, `docs/adr/0009-provenance-frontmatter-per-observation.md`, `docs/adr/0010-raw-transcripts-preserved-indefinitely.md`, `docs/adr/0011-coexistence-with-anthropic-auto-memory.md`

宿主 AI 硬性规则：
- 没有 repo_clone_verified=true 时，不得声称已经读过源码。
- 没有 repo_inspection_verified=true 时，不得把 README/docs/package 文件判断写成事实。
- 没有 quick_start_verified=true 时，不得声称 Quick Start 已跑通。

## Doramagic Pitfall Constraints / 踩坑约束

这些规则来自 Doramagic 发现、验证或编译过程中的项目专属坑点。宿主 AI 必须把它们当作工作约束，而不是普通说明文字。

### Constraint 1: 可能修改宿主 AI 配置

- Trigger: 项目面向 Claude/Cursor/Codex/Gemini/OpenCode 等宿主，或安装命令涉及用户配置目录。
- Host AI rule: 列出会写入的配置文件、目录和卸载/回滚步骤。
- Why it matters: 安装可能改变本机 AI 工具行为，用户需要知道写入位置和回滚方法。
- Evidence: capability.host_targets | https://github.com/LH8PPL/claude-memory-kit | host_targets=claude_code, claude
- Hard boundary: 不要把这个坑点包装成已解决、已验证或可忽略，除非后续验证证据明确证明它已经关闭。

### Constraint 2: 能力判断依赖假设

- Trigger: README/documentation is current enough for a first validation pass.
- Host AI rule: 将假设转成下游验证清单。
- Why it matters: 假设不成立时，用户拿不到承诺的能力。
- Evidence: capability.assumptions | https://github.com/LH8PPL/claude-memory-kit | README/documentation is current enough for a first validation pass.
- Hard boundary: 不要把这个坑点包装成已解决、已验证或可忽略，除非后续验证证据明确证明它已经关闭。

### Constraint 3: 维护活跃度未知

- Trigger: 未记录 last_activity_observed。
- Host AI rule: 补 GitHub 最近 commit、release、issue/PR 响应信号。
- Why it matters: 新项目、停更项目和活跃项目会被混在一起，推荐信任度下降。
- Evidence: evidence.maintainer_signals | https://github.com/LH8PPL/claude-memory-kit | last_activity_observed missing
- Hard boundary: 不要把这个坑点包装成已解决、已验证或可忽略，除非后续验证证据明确证明它已经关闭。

- Trigger: no_demo
- Evidence: downstream_validation.risk_items | https://github.com/LH8PPL/claude-memory-kit | no_demo; severity=medium
- Hard boundary: 不要把这个坑点包装成已解决、已验证或可忽略，除非后续验证证据明确证明它已经关闭。

### Constraint 5: 存在评分风险

- Trigger: no_demo
- Why it matters: 风险会影响是否适合普通用户安装。
- Evidence: risks.scoring_risks | https://github.com/LH8PPL/claude-memory-kit | no_demo; severity=medium
- Hard boundary: 不要把这个坑点包装成已解决、已验证或可忽略，除非后续验证证据明确证明它已经关闭。

### Constraint 6: issue/PR 响应质量未知

- Trigger: issue_or_pr_quality=unknown。
- Host AI rule: 抽样最近 issue/PR，判断是否长期无人处理。
- Why it matters: 用户无法判断遇到问题后是否有人维护。
- Evidence: evidence.maintainer_signals | https://github.com/LH8PPL/claude-memory-kit | issue_or_pr_quality=unknown
- Hard boundary: 不要把这个坑点包装成已解决、已验证或可忽略，除非后续验证证据明确证明它已经关闭。

### Constraint 7: 发布节奏不明确

- Trigger: release_recency=unknown。
- Host AI rule: 确认最近 release/tag 和 README 安装命令是否一致。
- Why it matters: 安装命令和文档可能落后于代码，用户踩坑概率升高。
- Evidence: evidence.maintainer_signals | https://github.com/LH8PPL/claude-memory-kit | release_recency=unknown
- Hard boundary: 不要把这个坑点包装成已解决、已验证或可忽略，除非后续验证证据明确证明它已经关闭。