Junyi Archive

自动归档 (Junyi Archive)

A three-layer knowledge management system for families. Turns the flood of daily information into organized, retrievable, and reusable wisdom.

Design Principles

• Low-friction capture, traceable distillation, cautious upgrades, long-term maintainability
• Only text content; images/video out of scope
• Storage: Obsidian physical files (not memory — survives Gateway restarts)

The Three Layers

Inbox (收集层)        → Raw capture, organized by scene
      ↓ AI distills (when warranted)
Cognition (认知层)    → Refined insights by 5 dimensions
      ↓ Proven patterns emerge
Guidebook (指南层)    → Verified reusable guides, methods, checklists

Layer 1: Inbox (收集层)

Zero-friction entry point. Raw records organized by scene.

Default scenes:

Custom scenes: When user proposes a new scene, confirm name once, then auto-create directory.

Layer 2: Cognition (认知层)

Distilled knowledge organized by five dimensions:

IMPORTANT: Use exact directory names above (with spaces, title case). Never convert to lowercase-hyphen format like what-i-think.

Classification rules:

• External source primarily → What I Learn
• Self-observation/reflection primarily → What I Think
• Finished output → What I Create
• Fragment to keep as-is → What I Collect
• Family identity/values/rules → Who We Are
• Mixed content → classify by primary source; if both parts are important, split into two cognition entries

Layer 3: Guidebook (指南层)

Proven, reusable methods that answer "how to do X" concretely.

Admission criteria (strict):

• User explicitly says "已验证", "反复有效", "用了较长时间", or similar
• OR original text contains clear validation evidence (e.g., "执行两个月有效")
• Methods that don't meet validation criteria stay in Cognition, not Guidebook
• Mark as "待验证" only if user wants to track an unvalidated method

Decision Flow (4 Steps)

Step 1: Determine user intent

• User says "就存一下 / 先记下来 / 不用总结" → Only write inbox, stop
• Normal record → Write inbox, then evaluate distillation (Step 3)
• Explicitly validated method/process → Write inbox + cognition + evaluate guidebook

Step 2: Determine scene

• Scene is clear → Route to corresponding inbox/ directory
• Scene is ambiguous but close → Pick closest match, preserve original context in content
• Scene truly unclear → Ask user, do not guess

Step 3: Determine whether to distill to Cognition

Distill when:

• Content has clear viewpoints, patterns, insights, or conclusions
• Content has long-term reuse value
• Content is a quote/fragment worth collecting
• Content is from articles/lectures/courses that can be distilled

Do NOT distill when:

• Pure memo or reminder
• Pure diary/stream-of-consciousness with no extractable insight
• User explicitly asked not to summarize
• Content only suitable for raw preservation

Short content exception: Short reflections (<200 chars) or single quotes can skip inbox and go directly to Cognition — the raw text IS the refined version, no need for duplicate storage.

Step 4: Determine whether to enter Guidebook

• Has executable steps + validated effectiveness → Guidebook
• Just an idea, principle, or feeling → Stay in Cognition
• Unvalidated method → Stay in Cognition, upgrade later when validated

File Management Rules

When to create new file

• New topic
• First record of the day under a new topic
• Content significantly different from existing notes
• Content belongs to a different layer or dimension

When to merge into existing file

• Same day AND same scene/topic
• New content supplements existing content (not a new topic)
• Append with a separator (---) and timestamp

When to update existing file

• Guidebook: prefer updating original file over creating new ones; refresh "更新日期"
• Cognition: if supplementing the same insight, add a "补充观察" section
• Never overwrite raw inbox content for cosmetic reasons

Naming Rules

• Inbox: YYYY-MM-DD 标题.md
• Cognition: 标题.md (no date prefix — cognition is topic-centric)
• Guidebook: 主题名.md

Title generation:

• If user provided a title, use it
• Otherwise generate a natural-language title from the first 8-15 chars of content
• Never use abstract titles like "感悟1" or "笔记2"

Content Templates

Inbox

# [Title]

> 日期：YYYY-MM-DD
> 场景：[scene name]
> 来源：[source if applicable]

---

[Original content — keep close to original, minimal formatting]

Cognition

# [Title]

> 日期：YYYY-MM-DD
> 维度：[dimension name]
> 来源：[source — article name, conversation, experience, etc.]
> 源文件：[corresponding inbox filename, e.g., inbox/学习成长/2026-03-13 蒙特梭利.md]
> 标签：#tag1 #tag2

---

[Refined content — distilled, structured, insightful]

## 核心要点
- Point 1
- Point 2

Note: When content skips inbox (short reflections/quotes), 源文件 field is "用户即时输入".

Guidebook

# [Title]

> 更新日期：YYYY-MM-DD
> 适用场景：[when to use this guide]
> 验证状态：[已验证 / 待验证]
> 来源：[corresponding cognition or inbox filename]

---

[Step-by-step guide, checklist, or methodology]

Field defaults:

• 来源 missing → "用户即时输入"
• Tags: 2-3 max, prefer topic/person/method keywords
• Preserve original attribution for articles/lectures/books

Source Traceability

• Every cognition entry must trace back to an inbox file or "用户即时输入"
• Every guidebook entry must trace back to cognition or inbox
• No orphaned conclusions without traceable context

Long Text Handling

Never use obsidian-cli create --content — truncation bug.

After every write, verify:

ls -la "<file>"     # exists and reasonable size
tail -5 "<file>"    # content is complete

Searching and Retrieval

• obsidian-cli search-content "<keyword>" — search note content
• obsidian-cli search "<note name>" — search note titles
• find + grep — precise filesystem search

Configuration

Vault path configured per user. Default structure (use exact directory names):

[vault root]/
├── inbox/
│   ├── 学习成长/
│   ├── 童言童语/
│   ├── 育儿观察/
│   ├── 灵感闪现/
│   ├── 对话收获/
│   ├── 工作记录/
│   └── 生活记录/
├── cognition/
│   ├── What I Learn/
│   ├── What I Think/
│   ├── What I Create/
│   ├── What I Collect/
│   └── Who We Are/
└── guidebook/