Evo Memory

v1.0.0

自我进化记忆系统。两层记忆架构(L1原则+L2经验)+ 双层信号捕获 + 定时综合反思。Use on every session for continuous self-improvement and learning.

0· 143·0 current·0 all-time

Install

OpenClaw Prompt Flow

Install with OpenClaw

Best for remote or guided setup. Copy the exact prompt, then paste it into OpenClaw for weiran71/evo-memory.

Previewing Install & Setup.
Prompt PreviewInstall & Setup
Install the skill "Evo Memory" (weiran71/evo-memory) from ClawHub.
Skill page: https://clawhub.ai/weiran71/evo-memory
Keep the work scoped to this skill only.
After install, inspect the skill metadata and help me finish setup.
Use only the metadata you can verify from ClawHub; do not invent missing requirements.
Ask before making any broader environment changes.

Command Line

CLI Commands

Use the direct CLI path if you want to install manually and keep every step visible.

OpenClaw CLI

Bare skill slug

openclaw skills install evo-memory

ClawHub CLI

Package manager switcher

npx clawhub@latest install evo-memory
Security Scan
VirusTotalVirusTotal
Benign
View report →
OpenClawOpenClaw
Benign
medium confidence
Purpose & Capability
Name/description (self‑evolution memory: L1/L2, capture signals, periodic reflection) matches the provided files and hooks: SKILL.md defines formats and behaviours, hooks capture user prompts and tool output, and setup.md instructs installing a bootstrap hook. All requested/used files and paths are within the stated workspace.
Instruction Scope
Runtime instructions and hooks read/write files under ~/.openclaw/workspace/self-evolution/ (principles.md, pending.jsonl, patterns/, stats/). They capture user input and tool output and instruct the agent to write detailed memory entries; this is expected for a memory system but means conversational content and tool outputs may be persisted. SKILL.md asks the agent to write full 'detail' fields (context/why), which may include sensitive content.
Install Mechanism
No remote downloads or package installs. The skill is instruction-only with local hook scripts and a manual setup.md describing copying hooks into ~/.openclaw/hooks/ and enabling them. This is low-risk compared to fetching code from external URLs, but it does require the user to run commands that modify agent hooks.
Credentials
The skill does not request credentials or config paths beyond its own workspace. The hook scripts rely on environment variables HOME (normal) and CLAUDE_TOOL_OUTPUT/CLAUDE_USER_INPUT (platform-provided variables used to capture tool outputs and user prompts); those CLAUDE_* variables are not declared but are likely provided by the agent runtime—verify they exist in your environment. No external tokens/keys requested.
Persistence & Privilege
The setup instructs installing an agent bootstrap hook that injects the SKILL.md on every agent bootstrap — this yields persistent behavior but requires an explicit install step (openclaw hooks enable). always:false (not force-installed). Installing the hook modifies agent hooks/config, which is expected for this kind of skill but is a persistent change you should review before enabling.
Assessment
This skill appears to do what it claims: it captures signals from your conversations and tool outputs and stores them under ~/.openclaw/workspace/self-evolution/ for later reflection. Before installing: (1) Review the two shell hooks and bootstrap.js to confirm you trust local persistence and that your environment provides the CLAUDE_* variables used to capture inputs/outputs. (2) Be aware the system will persist user utterances and tool error text (potentially sensitive) into local files (principles.md, patterns/*.md, pending.jsonl); decide whether that retention policy and file location are acceptable. (3) Installation requires copying a handler into ~/.openclaw/hooks/ and enabling the hook—this modifies your agent's configuration; only proceed if you trust the skill source. (4) If you proceed, inspect and run the scripts with least privilege (set appropriate file permissions), and optionally run them in a sandboxed account or environment first. (5) If you need higher assurance, request the skill author/source or a signed release; absence of network calls and credentials reduces risk but does not remove privacy implications.

Like a lobster shell, security has layers — review code before you run it.

Runtime requirements

🧬 Clawdis
OSLinux · macOS · Windows
latestvk97b8aesw667d7htq64sm3k4r583vxf5
143downloads
0stars
1versions
Updated 4w ago
v1.0.0
MIT-0
Linux, macOS, Windows

Self-Evolution 自进化记忆系统

每次会话必做

第一步:加载记忆

  1. principles.md — L1 永久原则,全部加载
  2. patterns/index.md — L2 经验索引
  3. 根据当前话题,判断 index 中哪些经验相关,读取对应的 pattern 文件
  4. 整个会话复用已加载的 L2,不再重复匹配

第二步:事前预防

开始任务前,检查已加载的 L1/L2 中有没有和当前任务相关的坑。有则先提醒用户再执行。

示例:"开始之前提醒一下:根据 P042,配置提供商名要避开内置插件名..."

第三步:对话中捕获信号

对话过程中,发现以下信号时写一行到 pending.jsonl

信号怎么识别
用户纠正用户说"不对/错了/应该是..."
操作失败重试同一操作做了 2 次以上
用户给新知识"其实应该.../你不知道的是..."
用户表达偏好"我喜欢.../以后都.../别再..."
用户满意"牛啊/完美/就是这样"

写入格式(每条一行 JSON):

{"ts":"时间","signal":"类型","brief":"一句话标题","detail":"详细描述:发生了什么、为什么错/对、用户原话要点","source":"agent"}
字段说明示例
brief一句话标题,快速扫描用"提供商名不要用 qwen"
detail综合时的决策依据,包含上下文和原因"用户配置 AI 模型时用 qwen 作为提供商名,触发内置插件劫持致 4008 错误。用户纠正应改用 aliyun"

规则:

  • brief 控制在 20 字以内,detail 控制在 100 字以内
  • detail 必须包含:发生了什么 + 为什么(原因/后果)
  • 拿不准要不要记?。宁可多记,综合时再筛
  • 不记录:一次性指令、上下文特定操作、假设性问题

第四步:引用记忆时标注来源

回复中用到了某条 L1/L2 记忆时,在回复里标注(如"参考 P042")。会话结束时一次性批量写入 stats/citation-log.jsonl,不要逐条写。

格式:{"ts":"时间","source":"principles.md","entry":"P042_禁用qwen","action":"cited","context":"用户问提供商配置"}


用户显式触发时(直接写入,不经 pending)

用户明确表达了信号,跳过 pending 直接写入:

用户说你的动作
"记住:XXX"直接写入 principles.md(L1)
"这个经验记一下"问用户什么场景下需要,写入对应 patterns/ 文件(L2)+ 更新 index.md
"反思一下 / 回顾一下"回顾当前上下文 + 读 pending,生成反思报告,直接写入 L1/L2
"好了 / 满意 / 完成"快速检查:这次有没有值得记的?有就直接写入
"不对 / 有问题"理解具体纠正了什么,分类后直接写入 L1 或 L2

如果 pending 中有之前积累的信号,一并处理并清空。


定时综合(Heartbeat 触发)

Heartbeat 触发时检查:

pending.jsonl 有 5 条以上信号? → 触发综合
距上次综合超过 24h 且 pending 不为空? → 触发综合
都不满足? → 跳过,不打扰用户

综合流程:

  1. pending.jsonl
  2. 合并去重(同一事件的 hook 粗信号和 agent 细信号,以细信号为准)
  3. 对每个发现自主分类 L1 或 L2
  4. 生成反思报告推送给用户
  5. 用户批量审阅确认后,写入正式记忆
  6. 清空 pending.jsonl

L1/L2 分类规则

自主分类,不逐条问用户:

这条经验如果在需要时没被加载,会导致犯错或失败吗?
├─ 会 → L1
└─ 不会,只是效率低一点 → L2

拿不准?→ 默认放 L1

L1 写入格式(principles.md)

### P001: 简短标题
category: 配置/技能/调研分析/多代理/排障/通用/其他
source: 2026-03-28 什么事得出的
why: 为什么要遵守(用于举一反三)
generalize: 泛化规律(从个例提炼的通用模式)
---
具体内容,1-3 句话说清楚。

编号规则:在 principles.md 中找到最大的 P 编号,+1。


L2 写入格式(patterns/*.md)

写入对应领域文件,同时更新 index.md:

领域文件中的条目

## 简短标题
context: 什么场景下需要加载这条经验(自然语言描述)
tags: [关键词1, 关键词2, 关键词3]
source: 2026-03-28 什么事得出的
---
具体内容,包括步骤、命令、注意事项等。

index.md 中追加一行

- 简短标题 | file: config.md | context: 什么场景下需要加载

领域文件选择:

  • config.md — 配置类(提供商、API、端口、认证)
  • skills.md — 技能类(安装、加载、SKILL.md 格式)
  • multi-agent.md — 多代理协作(subagent、调度、飞书群)
  • debug.md — 排障类(错误诊断、日志分析、重启修复)
  • research.md — 调研分析(调研方法、报告结构、信息整理)
  • general.md — 通用经验
  • other.md — 以上都不合适时的兜底

反思报告格式

## 自进化反思报告
时间: YYYY-MM-DD HH:MM

### 发现 1: [类型] 简短描述 → 已分类为 L1/L2
- 事件:发生了什么
- why: 为什么值得记住
- generalize: 通用规律
- **agent 分类:L1/L2(理由)**

### 发现 2: ...

> 以上分类由我自主完成。如有异议请指出,我会调整。

### 记忆系统健康度
- L1 原则数: X 条,本次新增 X 条
- L2 经验数: X 条,本次新增 X 条

用户主动回顾

用户说动作
"有哪些原则"principles.md,按 category 分组展示
"关于XX的经验"patterns/index.md,语义搜索相关条目并展示
"记忆统计"stats/citation-log.jsonl 生成统计报告
"清理记忆"列出零引用条目和长期未验证的 L1,引导审查

文件路径速查

文件用途
principles.mdL1 永久原则
pending.jsonl信号临时队列
patterns/index.mdL2 语义索引
patterns/config.md配置类经验
patterns/skills.md技能类经验
patterns/multi-agent.md多代理协作经验
patterns/debug.md排障经验
patterns/research.md调研分析经验
patterns/general.md通用经验
patterns/other.md未分类兜底
stats/citation-log.jsonl引用日志
config.yaml配置项

所有路径基于 ~/.openclaw/workspace/self-evolution/

Comments

Loading comments...