PARA Second Brain
Your agent's memory just got a massive upgrade. Full semantic search across your entire knowledge base — not just MEMORY.md.
What's New in v2.0
Before v2.0: memory_search only found content in MEMORY.md and daily logs. Your entire notes/ folder was invisible to search. You had to manually know where to look.
After v2.0: One symlink command makes your entire PARA knowledge base searchable. Ask about anything in your notes — it finds it. Plus session transcripts and memory flush protocol to prevent context loss.
| Before | After |
|---|
| Search only MEMORY.md + daily logs | Search EVERYTHING |
| "I don't have that information" |
Finds it instantly |
| Context compaction = lost information | Flush protocol saves critical context |
| Conversations forgotten | Session transcripts indexed |
What This Does
Creates a "second brain" structure that separates:
- - Raw capture (daily logs) from curated knowledge (MEMORY.md)
- Active work (projects) from ongoing responsibilities (areas)
- Reference material (resources) from completed work (archive)
How This Differs from Other Second Brain Skills
There's another popular second-brain skill powered by Ensue. Both are great — they solve different problems:
| PARA Second Brain (this skill) | Ensue Second Brain |
|---|
| Storage | Local files in your workspace | Cloud API (Ensue) |
| Cost |
Free, self-hosted | Requires Ensue API key |
|
Best for | Work context, agent continuity, project tracking | Evergreen knowledge base, semantic queries |
|
Search | Clawdbot's
memory_search | Ensue's vector search |
|
Structure | PARA (Projects/Areas/Resources/Archive) | Namespaces (concepts/toolbox/patterns) |
|
Use case | "What did we decide yesterday?" | "How does recursion work?" |
Use this skill if: You want file-based memory that works offline, costs nothing, and tracks ongoing work context.
Use Ensue's skill if: You want a cloud-hosted knowledge base optimized for semantic "what do I know about X" queries.
Use both if: You want PARA for work context + Ensue for evergreen knowledge. They complement each other.
Quick Setup
1. Create Directory Structure
CODEBLOCK0
Run this to scaffold:
CODEBLOCK1
2. Make Notes Searchable (The Symlink Trick)
By default, memory_search only indexes MEMORY.md and memory/*.md. Your entire notes/ folder is invisible to semantic search!
Fix this with one command:
CODEBLOCK2
Example:
CODEBLOCK3
What this does: Creates a symbolic link so memory/notes/ points to your actual notes/ folder. Now Clawdbot's memory_search sees all your PARA notes.
Verify it worked:
CODEBLOCK4
Test the search:
Ask your agent something that's in your notes but NOT in MEMORY.md. If it finds it, the symlink is working.
Why this matters:
| Before | After |
|---|
| Search only finds MEMORY.md + daily logs | Search finds ALL your notes |
| Must manually know where to look |
Semantic search across everything |
| "I don't have that information" | Finds connections you forgot existed |
3. Enable Session Transcript Indexing
Make your past conversations searchable too. Add this to your Clawdbot config:
CODEBLOCK5
What this does: Indexes your conversation transcripts alongside your notes. Now when you ask "what did we discuss about X last week?" — it can actually find it.
4. Initialize MEMORY.md
Create MEMORY.md in workspace root - this is your curated long-term memory:
CODEBLOCK6
5. Add to AGENTS.md
Add these instructions to your AGENTS.md:
CODEBLOCK7
Memory Flush Protocol (Critical!)
Your agent's context window is finite. When it fills up, older context gets compacted or lost. Don't lose important information.
How to Monitor
Run
session_status periodically. Look for:
CODEBLOCK8
Threshold-Based Actions
| Context % | What to Do |
|---|
| < 50% | Normal operation. Write decisions as they happen. |
| 50-70% |
Increased vigilance. Write key points after each substantial exchange. |
|
70-85% | Active flushing. Write everything important to daily notes NOW. |
|
> 85% | Emergency flush. Stop and write full context summary before responding. |
|
After compaction | Immediately note what context may have been lost. Check continuity. |
What to Flush
- 1. Decisions made — what was decided and why
- Action items — who's doing what
- Open threads — anything unfinished → INLINECODE11
- Working changes — if you discussed changes to files, make them NOW
Memory Flush Checklist
Before a long session ends or context gets high:
- - [ ] Key decisions documented?
- [ ] Action items captured?
- [ ] New learnings written to appropriate files?
- [ ] Open loops noted for follow-up?
- [ ] Could future-me continue this conversation from notes alone?
Knowledge Quality
The core question: "Will future-me thank me for this?"
What to Save
- - Concepts you actually understand (not half-learned ideas)
- Tools you've actually used (not just heard about)
- Patterns that worked (with concrete examples)
- Lessons learned from mistakes
What NOT to Save
- - Half-understood concepts (learn first, save after)
- Tools you haven't tried yet (bookmarks ≠ knowledge)
- Shallow entries without the WHY
- Duplicates of existing notes
Quality Gates
Before saving any curated note:
- 1. Written for future self who forgot context?
- Includes WHY, not just WHAT?
- Has concrete examples or key insight?
- Structured for retrieval (scannable)?
Content Templates
Use these for structured, high-quality entries in notes/resources/:
Concept Template
CODEBLOCK9
Tool Template
CODEBLOCK10
Pattern Template
CODEBLOCK11
PARA Explained
PARA is a knowledge organization system created by Tiago Forte, author of Building a Second Brain. It organizes everything into four categories based on actionability:
Projects
What: Work with a deadline or end state
Examples: "Launch website", "Plan trip to Japan", "Finish client proposal"
File as: INLINECODE13
Areas
What: Ongoing responsibilities with no end date
Examples: Health, finances, relationships, career development
File as: notes/areas/health.md, INLINECODE15
Resources
What: Reference material for future use
Examples: Research, tutorials, templates, interesting articles
File as: notes/resources/tax-guide.md, INLINECODE17
Archive
What: Inactive items from the other categories
Examples: Completed projects, outdated resources, paused areas
Move to: notes/archive/ when done
Daily Log Format
Create memory/YYYY-MM-DD.md for each day:
CODEBLOCK12
The Curation Workflow
Daily (5 min)
- - Log notable events to INLINECODE20
- File topic-specific notes to appropriate
notes/ folder
Weekly (15 min)
- - Review the week's daily logs
- Extract patterns and learnings to MEMORY.md
- Move completed projects to archive
Monthly (30 min)
- - Review MEMORY.md for outdated info
- Consolidate or archive old project notes
- Ensure areas reflect current priorities
Decision Tree: Where Does This Go?
CODEBLOCK13
Why Two Memory Layers?
| Daily Logs | MEMORY.md |
|---|
| Raw, timestamped | Curated, organized |
| Everything captured |
Only what matters |
| Chronological | Topical |
| High volume | Condensed |
| "What happened" | "What I learned" |
Daily logs are your journal. MEMORY.md is your wisdom.
Principles
- 1. Quality over quantity — Curated notes beat note hoarding
- Capture fast, curate deliberately — Daily logs are loose; curated notes are high quality
- Text > brain — If it matters, write it down
- Future-me test — "Will future-me thank me for this?"
- One home per item — Don't duplicate; link instead
- Include the WHY — Facts without context are useless
- Flush before you lose — Monitor context, write before compaction
Pairs well with memory-setup for technical config and proactive-agent for behavioral patterns.技能名称: para-second-brain
详细描述:
PARA 第二大脑
你的智能体记忆系统刚刚获得重大升级。现在支持对整个知识库进行完整的语义搜索——而不仅仅是 MEMORY.md 文件。
v2.0 新特性
v2.0 之前： memory_search 只能搜索 MEMORY.md 和每日日志中的内容。你整个 notes/ 文件夹对搜索来说都是不可见的。你必须手动知道去哪里查找。
v2.0 之后： 一个符号链接命令就能让你的整个 PARA 知识库变得可搜索。询问笔记中的任何内容——它都能找到。此外，还增加了会话记录索引和内存刷新协议，以防止上下文丢失。
| 之前 | 之后 |
|---|
| 仅搜索 MEMORY.md + 每日日志 | 搜索所有内容 |
| 我没有那个信息 |
立即找到 |
| 上下文压缩 = 信息丢失 | 刷新协议保存关键上下文 |
| 对话被遗忘 | 会话记录被索引 |
功能说明
创建一个第二大脑结构，用于分离：
- - 原始捕获（每日日志）与 精炼知识（MEMORY.md）
- 主动工作（项目）与 持续责任（领域）
- 参考资料（资源）与 已完成工作（归档）
与其他第二大脑技能的区别
还有一个由 Ensue 驱动的流行 第二大脑技能。两者都很棒——它们解决不同的问题：
| PARA 第二大脑（本技能） | Ensue 第二大脑 |
|---|
| 存储 | 工作区中的本地文件 | 云 API（Ensue） |
| 成本 |
免费，自托管 | 需要 Ensue API 密钥 |
|
最适合 | 工作上下文、智能体连续性、项目跟踪 | 常青知识库、语义查询 |
|
搜索 | Clawdbot 的 memory_search | Ensue 的向量搜索 |
|
结构 | PARA（项目/领域/资源/归档） | 命名空间（概念/工具箱/模式） |
|
用例 | 我们昨天决定了什么？ | 递归是如何工作的？ |
使用本技能，如果： 你想要基于文件的记忆系统，可以离线工作，零成本，并跟踪正在进行的工作上下文。
使用 Ensue 的技能，如果： 你想要一个针对语义我知道关于 X 的什么查询进行优化的云托管知识库。
两者都使用，如果： 你想要 PARA 用于工作上下文 + Ensue 用于常青知识。它们相辅相成。
快速设置
1. 创建目录结构
workspace/
├── MEMORY.md # 精炼的长期记忆
├── memory/
│ └── YYYY-MM-DD.md # 每日原始日志
└── notes/
├── projects/ # 有截止日期的主动工作
├── areas/ # 持续的生活责任
├── resources/ # 参考资料
│ └── templates/ # 内容模板
└── archive/ # 已完成/非活跃项目
运行此命令来搭建结构：
bash
mkdir -p memory notes/projects notes/areas notes/resources/templates notes/archive
2. 使笔记可搜索（符号链接技巧）
默认情况下，memory_search 只索引 MEMORY.md 和 memory/*.md。你整个 notes/ 文件夹对语义搜索来说是不可见的！
用一个命令修复：
bash
ln -s /path/to/your/workspace/notes /path/to/your/workspace/memory/notes
示例：
bash
ln -s /Users/yourname/clawd/notes /Users/yourname/clawd/memory/notes
作用： 创建一个符号链接，使 memory/notes/ 指向你实际的 notes/ 文件夹。现在 Clawdbot 的 memory_search 可以看到你所有的 PARA 笔记。
验证是否成功：
bash
ls -la memory/notes # 应显示：memory/notes -> /path/to/notes
测试搜索：
向你的智能体询问一个在你的笔记中但不在 MEMORY.md 中的内容。如果它能找到，说明符号链接生效了。
为什么这很重要：
| 之前 | 之后 |
|---|
| 搜索仅能找到 MEMORY.md + 每日日志 | 搜索能找到你所有的笔记 |
| 必须手动知道去哪里查找 |
语义搜索覆盖所有内容 |
| 我没有那个信息 | 找到你忘记存在的关联 |
3. 启用会话记录索引
让你过去的对话也变得可搜索。将此添加到你的 Clawdbot 配置中：
json
memorySearch: {
sources: [memory, sessions],
query: {
minScore: 0.3,
maxResults: 20
}
}
作用： 将你的对话记录与笔记一起索引。现在当你问我们上周关于 X 讨论了什么？——它真的能找到。
4. 初始化 MEMORY.md
在工作区根目录创建 MEMORY.md——这是你精炼的长期记忆：
markdown
MEMORY.md — 长期记忆
关于 [人类姓名]
活跃上下文
- - 当前关注领域
- 进行中的项目（摘要，非细节）
- 截止日期或时效性事项
偏好与模式
经验教训
关键日期
5. 添加到 AGENTS.md
将这些指令添加到你的 AGENTS.md：
markdown
记忆
每次会话你都是全新开始。这些文件是你的连续性：
- - 每日笔记： memory/YYYY-MM-DD.md — 发生事件的原始日志
- 长期记忆： MEMORY.md — 精炼的记忆（类似于人类的长期记忆）
- 主题笔记： notes/ — 按 PARA 结构组织（全部可通过 memory_search 搜索）
写入规则
- - 如果它有未来价值，立即写下来
- 不要依赖脑内笔记——它们无法在重启后存活
- 文字 > 大脑 📝
PARA 结构
- - 项目 (notes/projects/) — 有截止日期的主动工作
- 领域 (notes/areas/) — 持续责任（健康、财务、人际关系）
- 资源 (notes/resources/) — 参考资料、操作指南、研究
- 归档 (notes/archive/) — 已完成或非活跃项目
内存刷新协议
使用 session_status 监控你的上下文使用情况。在压缩清除你的记忆之前，将重要上下文刷新到文件中：
在实质性交流后记录关键点 |
| 70-85% | 主动刷新——立即写下所有重要内容 |
| > 85% | 紧急刷新——在下一次响应前完成完整摘要 |
| 压缩后 | 记录可能丢失的上下文 |
规则： 根据阈值行动，而不是凭感觉。如果重要，立即写下来。
内存刷新协议（关键！）
你的智能体的上下文窗口是有限的。当它填满时，较旧的上下文会被压缩或丢失。不要丢失重要信息。
如何监控
定期运行 session_status。查找：
📚 上下文：36k/200k (18%) · 🧹 压缩次数：0
基于阈值的操作
| 上下文百分比 | 做什么 |
|---|
| < 50% | 正常运行。边发生边记录决策。 |
| 50-70% |
提高警惕。每次实质性交流后记录关键点。 |
|
70-85% | 主动刷新。立即将所有重要内容写入每日笔记。 |
|
> 85% | 紧急刷新。停止并在响应前写下完整的上下文摘要。 |
|
压缩后 | 立即记录可能丢失的上下文。检查连续性。 |
需要刷新什么
- 1. 做出的决策 — 决定了什么以及原因
- 行动项 — 谁在做什么
- 未完成事项 — 任何未完成的内容 → notes/areas/open-loops.md
- 工作中的更改 — 如果你讨论了文件更改，立即执行
内存刷新检查清单
在长时间会话结束或上下文变高之前：
- - [ ] 关键决策已记录？
- [ ] 行动项已捕获？
- [ ] 新学到的知识已写入相应文件？
- [ ] 未完成事项已记录以便跟进
