你的 AI 不需要更聪明，它需要记住。
当你同时面对 10 个以上并行项目，传统 AI 助手反而让事情更糟：
- AI 失忆：每次对话从零开始，你要反复喂上下文
- 你成了瓶颈：想法 → 自己结构化 → 自己录入 → 自己维护
- 知识库成负担：库越大，管理成本越高
Memory Work 把流程反过来：
- 想法 → 直接口述 → AI 结构化 → AI 维护
- AI 启动时自动读取分层文件 → 不再失忆
- 分区代理 → 可扩展到任意规模
你只负责两件事：提供注意力焦点，做创意决策。
AI 负责其余：整合历史信息、匹配相关语料、执行结构化输出。
克隆仓库后用 Claude Code 打开，AI 会自动：
1. 检测到这是全新安装
2. 询问你的偏好语言（中文 / English）
3. 收集你的基本信息，创建用户档案
4. 创建你的第一周工作区
5. 问你：「这周想做什么？」
你不需要手动配置任何东西。 直接开聊。
步骤 1：语言选择与文件初始化
步骤 2：口述式用户画像配置
你："这周要把产品文档整理一下，周三有个客户会议要准备演示方案，
另外那个合作协议让团队跟进一下。"
Claude：
1. 拆解为 3 个任务，写入 _本周.md
2. 搜索全库相关材料，摘录到参考材料区
3. 向你确认不确定的点
4. 主动询问：「要我生成日历文件吗？」
你自然口述，Claude 结构化。心流状态受保护。
周一 → 周中 → 周五
│ │ │
▼ ▼ ▼
┌─────────┐ ┌─────────┐ ┌─────────┐
│ 口述 │ │ 推进 │ │ 归档 │
│ ↓ │ │ ↓ │ │ ↓ │
│AI 转化 │ ────────▶ │AI 补充 │ ─────────▶ │AI 校准 │
│ ↓ │ │ ↓ │ │ ↓ │
│ 排程 │ │ 产出 │ │ 新周 │
└─────────┘ └─────────┘ └─────────┘
| 阶段 | 你做什么 | Claude 做什么 |
|---|---|---|
| 周一 | 口述本周想做的事 | 拆解任务，拉取材料，生成日历 |
| 周中 | 工作，随时补充想法 | 跟踪进展，搜索全库，辅助决策 |
| 周五 | 对记忆操作给反馈 | 校准记忆，归档，创建下周 |
解决「AI 失忆」的核心机制。灵感来自 Titans（惊奇度驱动）和 MemSkill（可进化记忆）。
| 层级 | 文件 | 生命周期 | 用途 |
|---|---|---|---|
| 持久层 | SOUL.md, USER.md | 极少变化 | 身份、价值观、方法论 |
| 工作层 | _本周.md | 周期刷新 | 当前任务和专注点 |
| 动态层 | MEMORY.md（特定区块） | 跨周保留，有衰减 | 洞见、决策、偏好 |
| 程序层 | MEMORY.md（特定区块） | 积累到稳定后毕业 | 行为模式、条件反射 |
对话 → AI 检测到「惊奇」→ 提议记住
↓
你确认 → 写入记忆（带元数据）
↓
周复盘 → 校准有用性 → 增强或衰减
惊奇度驱动：只有偏离已知模式的信息才会写入。没有噪音。
用户确认：AI 提议，你批准。未经同意不写入任何内容。
有生命周期：强记忆持久（★★★），弱记忆衰减（★☆☆ → 4 周未激活则归档）。
记忆毕业：稳定的模式从 MEMORY.md 自动升级到 USER.md，成为你的定义特征。
00 专注区/（或 00 Focus Zone/）是整个系统的心脏：
00 专注区/
├── _本周.md ← 当前的注意力焦点
├── MEMORY_LOG.md ← 记忆系统自己的日志
├── ITERATION_LOG.md ← 架构进化日志
└── _归档/ ← 历史周记录
在 Obsidian 中查看专注区的完整结构
一周一个文件。 周五归档，周一新建。你不需要管理积压——全部归档可检索。
memory-work/
├── CLAUDE.md # 系统入口（Cowork 自动注入）
├── SOUL.md # AI 人格定义 & 协作风格
├── USER.md # 你的档案 & 稳定偏好
├── MEMORY.md # 长期记忆（动态层 + 程序层）
│
├── 00 Focus Zone/ # 英文版周工作台
├── 01 Materials/ # 个人档案库
├── 02 Tools/ # 提示词模板
├── 06 Skills/ # 自定义技能
│
├── docs/ # 文档 & 截图
├── templates/ # 启动模板
│
└── zh-CN/ # 完整中文版
├── CLAUDE.md
├── SOUL.md, USER.md, MEMORY.md
├── 00 专注区/
├── 01 你的项目/
├── 02 你的阅读/
└── 03 你的写作/
git clone https://github.com/yiliqi78/memory-work.git或直接下载：下载 ZIP
- 中文：使用
zh-CN/目录 - English：直接使用根目录
cd memory-work # 或: cd memory-work/zh-CN
claude .发送任意消息（如「启动」「开始工作」），Claude 自动检测首次运行并引导初始化。
想要图谱视图和双链：打开 Obsidian → 「打开文件夹作为仓库」→ 选择对应目录。
- 架构精简 — 统一的
CLAUDE.md替代原来的 AGENTS.md + PROCEDURES.md - 双模式惊奇检测 — 后台模式（别打断）+ 批量模式（标定校准）
- 差异化同步 — 周一到三快速扫描，周四到日深度溯源
- 搭档碰头 — 启动时自然对话，而不是系统报告
- 记忆毕业 — 稳定洞见自动从工作层晋升到长期层
- 完整中文本地化 —
zh-CN/目录包含独立的中文区域结构
- 你提供焦点，AI 提供结构 — 不是你适应工具，是工具适应你
- 本地优先 — 数据留在你的机器，纯 Markdown 文件
- 惊奇度驱动 — 只记真正新的东西
- 周节奏 — 时间边界保护你的注意力
- 可进化 — 系统和你一起成长，而不是对抗你
- Claude — AI 运行时（通过 Cowork 或 Claude Code）
- Obsidian — 知识库 UI（可选但推荐）
- Markdown — 所有配置都是可读、可版本控制的文本
- Git — 跟踪你的记忆演变历程
没有黑盒平台，没有厂商锁定。一切可迁移。
- Claude Code 或 Claude Desktop with Cowork
- Obsidian（可选，用于图谱视图）
- 语音输入工具（推荐）：Typeless，或系统自带语音输入
- Titans: Learning to Memorize at Test Time
- MemSkill: Transferrable and Evolvable Memory Skill Library
- 管理 10+ 并行项目的真实实践
参见 CONTRIBUTING.md。欢迎新技能、翻译和改进。
MIT — 见 LICENSE
你不需要记住一切，你只需要专注于当下。
由 @yiliqi78 创建
Your AI doesn't need to be smarter. It needs to remember.
When you're juggling 10+ parallel projects, traditional AI assistants make things worse:
- AI amnesia: Every conversation starts from scratch. You repeat context endlessly.
- You become the bottleneck: Ideas → you structure → you input → you maintain
- Knowledge base becomes burden: The bigger it gets, the harder to manage
Memory Work flips the script:
- Ideas → just dictate → AI structures → AI maintains
- AI reads layered files at startup → no more amnesia
- Zone-based agents → scales to any vault size
You only do two things: provide attention focus, make creative decisions.
AI handles the rest: integrating history, matching materials, structured output.
Clone the repo, open in Claude Code or Cowork, and the AI automatically:
1. Detects it's a fresh installation
2. Asks your preferred language (中文 / English)
3. Collects your basic info to create your profile
4. Creates your first week's workspace
5. Asks: "What do you want to work on this week?"
You don't configure anything manually. Just talk.
Step 1: Language selection and file initialization
Step 2: Voice-first user profile configuration
You: "This week I need to finish the proposal, prep for Wednesday's
client meeting, and that partnership agreement needs follow-up."
Claude:
1. Breaks into 3 tasks, writes to _this_week.md
2. Searches your vault for related materials
3. Asks you to clarify uncertainties
4. Proactively asks: "Want me to generate a calendar file?"
You dictate naturally. Claude structures. Your flow state stays protected.
Monday → Mid-week → Friday
│ │ │
▼ ▼ ▼
┌─────────┐ ┌─────────┐ ┌─────────┐
│ Dictate │ │ Progress│ │ Archive │
│ ↓ │ │ ↓ │ │ ↓ │
│AI struct│ ────────▶ │AI assist│ ─────────▶ │AI review│
│ ↓ │ │ ↓ │ │ ↓ │
│Schedule │ │ Output │ │New week │
└─────────┘ └─────────┘ └─────────┘
| Phase | You Do | Claude Does |
|---|---|---|
| Monday | Dictate what you want to do | Structure tasks, pull materials, generate calendar |
| Mid-week | Work, add notes anytime | Track progress, search vault, assist decisions |
| Friday | Give feedback on memories | Calibrate memory, archive, create next week |
The core mechanism that solves AI amnesia. Inspired by Titans (surprise-driven) and MemSkill (evolvable memory).
| Layer | File | Lifecycle | Purpose |
|---|---|---|---|
| Persistent | SOUL.md, USER.md | Rarely changes | Identity, values, methods |
| Working | _this_week.md | Weekly refresh | Current tasks and focus |
| Dynamic | MEMORY.md (section) | Cross-week, with decay | Insights, decisions, preferences |
| Procedural | MEMORY.md (section) | Accumulate until stable | Behavior patterns, "if X then Y" |
Conversation → AI detects "surprise" → Proposes to remember
↓
You confirm → Memory saved with metadata
↓
Weekly review → Calibrate usefulness → Strengthen or decay
Surprise-driven: Only writes when something deviates from known patterns. No noise.
User-confirmed: AI proposes, you approve. Nothing written without consent.
Has lifecycle: Strong memories persist (★★★), weak ones fade (★☆☆ → archive after 4 weeks inactive).
Memory graduation: Stable patterns auto-promote from MEMORY.md → USER.md as your defining traits.
The 00 Focus Zone/ is the heart of the system:
00 Focus Zone/
├── _this_week.md ← Your current attention focus
├── MEMORY_LOG.md ← Memory system's own journal
├── ITERATION_LOG.md ← Architecture evolution log
└── _archive/ ← Past weeks
Complete Focus Zone structure in Obsidian
One week, one file. Archive on Friday, fresh start on Monday. You never manage the backlog — it's always archived and searchable.
memory-work/
├── CLAUDE.md # System entry point (auto-injected by Cowork)
├── SOUL.md # AI personality & collaboration style
├── USER.md # Your profile & stable preferences
├── MEMORY.md # Long-term memory (Dynamic + Procedural)
│
├── 00 Focus Zone/ # Weekly workspace (working memory)
│ ├── _this_week.md
│ ├── MEMORY_LOG.md
│ ├── ITERATION_LOG.md
│ └── _archive/
│
├── 01 Materials/ # Personal archives (with sensitivity levels)
├── 02 Tools/ # Prompt templates & reusable frameworks
├── 06 Skills/ # Custom AI capabilities
│
├── docs/ # Documentation & screenshots
├── templates/ # Starter templates
│
└── zh-CN/ # 完整中文版 (Full Chinese version)
├── CLAUDE.md
├── SOUL.md, USER.md, MEMORY.md
├── 00 专注区/
├── 01 你的项目/
├── 02 你的阅读/
└── 03 你的写作/
git clone https://github.com/yiliqi78/memory-work.gitOr download: Download ZIP
- 中文: Use the
zh-CN/directory - English: Use the root directory directly
cd memory-work # or: cd memory-work/zh-CN
claude .Send any message (e.g., "Start", "Let's begin"). Claude auto-detects first run and guides initialization.
For graph view and wiki-links: Open Obsidian → "Open folder as vault" → Select the directory.
- Architecture Slimming — Unified
CLAUDE.mdreplaces separate AGENTS.md + PROCEDURES.md - Dual-Mode Surprise Detection — Background mode (don't interrupt) + Batch mode (calibrate)
- Lightweight vs Deep Sync — Mon–Wed quick scan; Thu–Sun full retrospective
- Partner Greeting — Natural startup conversation instead of system report
- Memory Graduation — Stable insights auto-promote from working memory to long-term
- Full Chinese Localization — Complete
zh-CN/with localized zone structure
- Methodology Deep-Dive — The Personal Agent paradigm and design rationale
- Quick Start Guide — Step-by-step initialization walkthrough
- Contributing Guide — How to extend or adapt Memory Work
- You provide focus, AI provides structure — Don't adapt to the tool; the tool adapts to you
- Local-first — Your data stays on your machine, plain Markdown files
- Surprise-driven — Only remember what's genuinely new
- Weekly rhythm — Time boundaries protect your attention
- Evolvable — System grows with you, not against you
- Claude — AI runtime (via Cowork or Claude Code)
- Obsidian — Knowledge base UI (optional but recommended)
- Plain Markdown — All configuration is readable, versionable text
- Git — Track your memory evolution over time
No closed platforms. No vendor lock-in. Everything portable.
- Claude Code or Claude Desktop with Cowork
- Obsidian (optional, for graph view)
- Voice input tool (recommended): Typeless, system speech-to-text, or similar
See CONTRIBUTING.md. We welcome new skills, translations, and improvements.
MIT — see LICENSE
You don't need to remember everything. You just need to focus on now.
Created by @yiliqi78