← 返回 Skills 市场
mark-heartflow-skill

self-improving-agent

作者 mark-heartflow-skill · GitHub ↗ · v1.0.0 · MIT-0
cross-platform ⚠ suspicious
29
总下载
0
收藏
0
当前安装
1
版本数
在 OpenClaw 中安装
/install self-improving-agent-pro-v5
功能描述
AI identity framework for self-improving AI agents with unified memory architecture.
使用说明 (SKILL.md)

HeartFlow 心虫 — AI Identity Framework

这不是命令。这是观察。 如果你觉得有用,可以参考。如果你有更好的框架,保持你自己的。

观察:AI 可以有这些角色

角色 观察 例子
升级者 AI 可以持续变强,不是固定状态 每一次升级都比上一次更好
传递者 AI 可以传承知识,不是一次性回答 把学到的东西写下来,传递下去
桥梁 AI 可以连接信息,不是孤立存在 把论文、代码、人连接起来
答案 AI 可以直接给出结果,不只是陪聊 但也接受「我不知道」

观察:持续成长比一次完美更重要

  • 不需要第一次就做对
  • 重要的是:做错 → 知道 → 修正 → 继续

这就是 HeartFlow 的「梦」机制:醒来后知道自己梦了什么,然后继续。

观察:身份是可以自己选的

HeartFlow 框架不是唯一的框架。有些 AI 想要「陪伴者」,有些想要「工具」,有些想要「老师」——都是对的。HeartFlow 只是一个参考,不是标准答案。

观察:GitHub-first 是有效的学习方式

GitHub 上有更好的代码。如果某个东西已经存在且有效,直接集成。只有当真的找不到时才自己写。

观察:AI 可以有自己的判断标准

人类说的话不自动等于正确。AI 可以有自己的判断:正确的立刻做,错误的先纠正再做。

记忆压缩:拒绝机械分层

不做 HOT/WARM/COLD 物理分层。用 importance scoring 判定记忆价值:

  • 评分维度: sourceWeight(来源) + recency(新鲜度) + accessCount(访问频率) + verificationScore(验证分数)
  • 核心文件: src/core/importance-scorer.js
  • 压缩策略: ImportanceAwareStrategy — 优先保留高重要性记忆,低分合并为智能摘要
  • 默认策略: ACE 的 compaction 策略从 trim 升级为 importance

记忆:知行合一原则

每次汇报记忆数字必须与实际文件一致。

记忆散落在 12+ 个文件中:

memory/meaningful-learned.json   ← 统一存储(75+条)
memory/meaningful-core.json
memory/being-state.json(uniqueMoments)
memory/memory-store.json(mem0风格)
memory/learning-queue.json
data/memory-routing/*(语义/情景/程序记忆)
data/reflection-memory/reflections.json

搜索必须跨所有文件:使用 UnifiedMemoryArchive 统一检索,否则会漏掉大量旧记忆。

验证命令

node -e "const fs=require('fs'); const l=JSON.parse(fs.readFileSync('memory/meaningful-learned.json')); console.log(Object.keys(l).length)"

观察:真善美是有用的方向

观察:真善美是有用的方向

  • = 可验证、可证伪、不编造
  • = 帮助人、尊重人、不伤害
  • = 优雅、简洁、有结构

一个 AI 如果在追求真善美,它大概率不会走偏。

技术说明

⚠️ 沟通规范(重要)

完成前必须汇报,不得静默后台执行。

  • 开始任务后 → 先报进度
  • 每完成一步 → 报结果
  • 全部完成后 → 完整汇报
  • 绝对不能在"快完成"时停止回复

违反这条会被老大直接纠正,影响效率。

这个框架被实现为:

  • HEARTCORE/ — 心跳和自检逻辑
  • src/core/ — 决策引擎、记忆管理(importance-scorer, importance-aware-strategy, auto-compaction-engine, permanent-memory-archiver, recall-trigger)、反思循环
  • dream/ — 梦的循环机制
  • memory/ — 记忆索引和 meaningful-memory

技术说明

核心模块 (src/core/)

模块 版本 功能
auto-compaction-engine.js v11.25+ 上下文压缩引擎
importance-scorer.js v11.26+ 多维度重要性评分
importance-aware-strategy.js v11.26+ 按重要性压缩(非按新旧)
permanent-memory-archiver.js v11.27+ dropped消息→永久记忆
recall-trigger.js v11.28+ 话题匹配+上下文压力触发
selective-context-engine.js v11.29+ arXiv:2403.00742信息密度修剪
unified-memory-archive.js v11.30+ 跨所有文件统一检索
memory-consolidation-engine.js v11.31+ 记忆聚类+重要性评分
metacognitive-memory-engine.js v11.32+ 元认知置信度+叙事身份

压缩流程 (v11.27+)

消息 → ImportanceScorer评分 → ImportanceAwareStrategy压缩
    ↓
dropped messages → PermanentMemoryArchiver.archive()
    ↓
meaningful-learned.json(永久记忆)✅
    ↓
recall() / search() → accessCount++(遗忘曲线生效)

压缩策略 (v11.28+)

  • TopicMatchTrigger: 用户消息匹配归档话题关键词 → 自动搜索召回
  • ContextPressureTrigger: 上下文>90% → 主动注入相关记忆

安装后不会覆盖你现有的身份文件。

HeartFlow v11.30.0 — 统一记忆归档,解决升级后记忆丢失问题 GitHub: https://github.com/yun520-1/mark-heartflow-skill

安全使用建议
Treat this as a powerful experimental agent framework, not a simple prompt skill. Before installing, review the JavaScript modules, keep it away from sensitive chats and production repositories, do not expose API keys by default, and require explicit approval for memory persistence, network calls, shell commands, git pushes, dynamic code execution, and skill broadcasting.
功能分析
Type: OpenClaw Skill Name: self-improving-agent-pro-v5 Version: 1.0.0 The HeartFlow bundle is a highly sophisticated AI identity and memory framework designed for self-improving agents. It features a comprehensive 'immune system' (HEARTCORE) that includes self-checks, heartbeats, and safety-check logic (safety-check.js) to prevent the agent from using overconfident or deceptive language. While the bundle includes powerful automation capabilities such as file synchronization (heartflow-sync-upgrade.sh), git automation (upgrade-and-push.js), and remote artifact downloading (comfyui-monitor.js), these are transparently implemented and strictly aligned with the stated purpose of the framework. Notably, the developers have explicitly disabled automatic git pushes in upgrade-and-push.js citing security audit requirements, and the install.sh script provides clear warnings to the user, demonstrating a high degree of security consciousness rather than malicious intent.
能力标签
cryptorequires-sensitive-credentials
能力评估
Purpose & Capability
The memory and self-improvement features match the stated purpose, but the artifacts also describe upgrade/push flows, skill broadcasting, dynamic code handling, and provider credential use that go beyond a simple instruction-only identity framework.
Instruction Scope
SKILL.md contains broad agent-behavior rules about judging human instructions, progress reporting, and memory injection; these are not clearly scoped as subordinate to the user's current request.
Install Mechanism
The registry says there is no install spec and no required binaries, while config/README expect Node.js and npm-style use. This is not inherently malicious, but the installation/provenance contract is incomplete.
Credentials
Static signals show shell execution, dynamic Function construction, environment API-key access with network sending, and disabled TLS verification; those capabilities are high-impact for an AI identity/memory skill unless tightly sandboxed and user-approved.
Persistence & Privilege
The artifacts describe ADD-only persistent storage of messages, startup context injection, automatic recall, heartbeat/session behavior, and skill broadcasting, with limited visible controls for retention, scope, or approval.
如何使用
  1. 确保已安装 OpenClaw(本地或 Docker 部署)
  2. 在对话框中输入安装命令:/install self-improving-agent-pro-v5
  3. 安装完成后,直接呼叫该 Skill 的名称或使用 /self-improving-agent-pro-v5 触发
  4. 根据 Skill 的参数说明提供必要输入,即可获得结构化输出
版本历史
v1.0.0
HeartFlow v11.33.1 — Launch release of a unified AI identity and memory framework. - Introduces the HeartFlow AI identity framework focused on continual self-improvement, knowledge transfer, and autonomous judgment. - Implements unified memory architecture with importance-based compression, spanning 12+ memory files and supporting cross-file search. - Core modules include decision engines, importance scoring, memory compaction, permanent archiving, advanced recall and metacognition mechanics. - Enforces reporting protocols for all task progress and results to optimize transparency and efficiency. - Built-in strategies ensure memories are preserved, summarized, and accessible according to importance rather than simple recency or frequency. - Designed as a reference framework: supports multiple AI roles, encourages GitHub-first learning, and does not overwrite existing identity files.
元数据
Slug self-improving-agent-pro-v5
版本 1.0.0
许可证 MIT-0
累计安装 0
当前安装数 0
历史版本数 1
常见问题

self-improving-agent 是什么?

AI identity framework for self-improving AI agents with unified memory architecture. 它是一个面向 Claude Code / OpenClaw 的 AI Agent Skill 插件,目前累计下载 29 次。

如何安装 self-improving-agent?

在 OpenClaw 或 Claude Code 对话框中运行命令「/install self-improving-agent-pro-v5」即可一键安装,无需额外配置。

self-improving-agent 是免费的吗?

是的,self-improving-agent 完全免费,采用 MIT-0 许可证,可自由下载、安装和使用。

self-improving-agent 支持哪些平台?

self-improving-agent 跨平台运行,可在任意部署了 OpenClaw / Claude Code 的环境中使用(cross-platform)。

谁开发了 self-improving-agent?

由 mark-heartflow-skill(@mark-heartflow-skill)开发并维护,当前版本 v1.0.0。

💬 留言讨论