An Experience-Driven Self-Improvement Framework for LLM Agents
基于经验的 LLM Agent 自我改进框架
Status / 状态: Design proposal — looking for implementers 当前状态: 设计阶段，寻求实现者
English
What it is
Hindsight is a design specification for giving LLM agents a missing capability: the ability to learn from their own mistakes across sessions, and eventually internalize those lessons into permanent behavior.
Today's agents are amnesiacs. They make the same mistake on Tuesday that they made on Monday. They get corrected, apologize, and forget. Hindsight proposes a layered framework where:
- Errors are captured as structured lessons with metadata
- Lessons are retrieved before similar tasks to prevent repetition
- High-frequency, well-validated lessons can be compiled into the agent's permanent behavior
- Positive patterns are tracked symmetrically in a separate library
- The system avoids self-delusion by relying on real error data only
Why it exists
Most "memory" features in agent platforms today are RAG over previous conversations. That's not learning — that's recall.
Real learning requires:
- A structured record of what went wrong (not just what was said)
- A retrieval mechanism that surfaces lessons at the right moment
- A way to distinguish lessons that should remain "reminders" from lessons that should be baked into the system itself
- Safeguards against ossification, self-delusion, and rule conflicts
- A symmetric record of what works well, not just what fails
No current agent platform — Coze, Hermes, LangChain, Dify, OpenClaw, or even Claude Code — provides this end-to-end. Hindsight is a proposal for what such a system should look like.
The vision
The endgame is an agent that:
- Day 1: Makes a mistake. User corrects it. The mistake is recorded as a structured lesson.
- Day 7: About to make the same mistake. The relevant lesson surfaces. The agent avoids it.
- Day 30: The lesson has been hit 20 times reliably. It gets compiled into the agent's system prompt or behavioral code.
- Day 31: The agent no longer needs to "remember" the lesson — the mistake is structurally impossible.
This is not "self-evolving AI" in the science-fiction sense. It is constraint learning + experience accumulation + selective compilation, applied as a coherent system. It is closer in spirit to Reflexion (Shinn et al., 2023) and Voyager (Wang et al., 2023) than to AGI fantasies.
Current status
This is a design specification, not a working implementation. The author is not a developer. The author is publishing this to find people who can build it.
The design is split into two layers in DESIGN.md:
- Part 1 — Vision Architecture: The complete, ambitious system. This is the north star.
- Part 2 — MVP v2.0: The minimal viable subset that can be built today, by one person, in a reasonable timeframe.
- Part 3 — Evolution Path: How to grow from MVP to the full vision.
If you want to:
- Discuss the design → open an Issue
- Implement an MVP → see
DESIGN.md, Part 2 - Build the full vision → see
DESIGN.md, Part 1 - Adapt it to a specific platform (LangChain, Coze, OpenClaw, etc.) → fork it, share what you learn
Who this is for
Hindsight is most useful for:
- Agent operators with repetitive workflows (legal, procurement, customer support, content production)
- Enterprise deployments where real error data is abundant
- Power users who run the same agent across many sessions
It is less useful for one-off consumer chats, where the lesson library never accumulates enough signal.
License
MIT. Take it, build it, ship it. Just credit the design.
Acknowledgments
This framework was conceived and designed by the author — a non-developer who identified a real gap in how LLM agents handle failure, and spent considerable time thinking through the full architecture: the lesson lifecycle, the pre-task retrieval mechanism, the blind-test evolution validation, the two-phase token-efficient loading, and the bipolar memory model (lessons + patterns). None of these came from AI assistants.
Claude (Anthropic) and GPT (OpenAI) served as sounding boards: stress-testing the design, pointing out engineering blind spots, and helping translate the author's intuitions into technical language. The AI assistants contributed critique and articulation — the ideas themselves belong to the author.
中文版
这是什么
Hindsight 是一份设计规范，目标是给 LLM Agent 补上一个关键能力：从自己的错误中学习，跨会话累积教训，最终把教训内化到永久行为中。
今天的 Agent 都是健忘症患者。它们周一犯的错，周二会一字不差地再犯一遍。被纠正，道歉，然后遗忘。Hindsight 提出一个分层框架：
- 错误被捕获为结构化教训（带元数据，不是纯文本）
- 在执行类似任务前主动检索相关教训，避免重复
- 高频且验证有效的教训，可以被编译进 Agent 的永久行为
- 正向经验在独立的 Pattern 库里对称记录
- 系统通过只采用真实错误数据来避免自我陶醉式优化
为什么需要
目前 Agent 平台的"记忆"功能，本质上都是对历史对话做 RAG。这不是学习，这是回忆。
真正的学习需要：
- 对"哪里错了"的结构化记录（而不只是"说过什么"）
- 在合适时机自动浮现相关教训的检索机制
- 区分"需要持续提醒的教训"和"应该固化进系统的教训"的机制
- 防僵化、防自嗨、防规则冲突的安全锁
- 对称的"什么是好答案"的正向经验记录
目前没有任何 Agent 平台——扣子、Hermes、LangChain、Dify、OpenClaw、甚至 Claude Code——端到端提供这些能力。Hindsight 是对这套系统应该长什么样的一份提案。
终极愿景
最终目标是一个这样的 Agent：
- 第 1 天：犯了一个错。用户纠正。错误被记录为结构化教训。
- 第 7 天：即将犯同样的错。相关教训自动浮现。Agent 避开了。
- 第 30 天：这个教训已经被可靠地命中 20 次。它被编译进 Agent 的 system prompt 或行为代码。
- 第 31 天：Agent 不再需要"记住"这个教训——这个错误从结构上就不可能再发生。
这不是科幻意义上的"AI 自我进化"。这是约束学习 + 经验累积 + 选择性编译的协同系统。在精神上，它更接近 Reflexion（Shinn 等，2023）和 Voyager（Wang 等，2023）这些研究，而不是 AGI 幻想。
当前状态
这是一份设计规范，不是工作中的实现。作者不是开发者。作者发布这份规范是为了找到能够实现它的人。
设计被分成两层，在 DESIGN.md 里：
- Part 1 — 愿景架构：完整、有野心的系统。这是北极星目标。
- Part 2 — MVP v2.0：今天就能由一个人在合理时间内做出来的最小可行子集。
- Part 3 — 演化路径：从 MVP 长到完整版的可能路线。
如果你想：
- 讨论设计 → 开 Issue
- 实现 MVP → 看
DESIGN.md的 Part 2 - 构建完整愿景 → 看
DESIGN.md的 Part 1 - 适配到具体平台（LangChain、扣子、OpenClaw 等）→ Fork 一份，分享你学到的东西
适合谁
Hindsight 最有价值的场景：
- 有重复性工作流的 Agent 运营者（法律、招投标、客服、内容生产）
- 企业部署场景，因为有充足的真实错误数据
- 在多个会话中长期使用同一个 Agent 的高频用户
不太适合一次性的消费级对话，因为教训库永远累积不到足够的信号量。
许可证
MIT。拿去用、做出来、发布出去。只需要在归属里署名设计来源。
致谢
这套框架由作者构思并设计——一位不会写代码的人，发现了 LLM Agent 处理错误的方式存在真实缺陷，并花了大量时间思考完整的架构：教训的生命周期、任务前检索机制、进化验证中的盲测设计、两阶段 token 节省加载、双极记忆模型（教训库 + 模式库）。这些核心想法没有一个来自 AI 助手。
Claude（Anthropic）和 GPT（OpenAI）充当了"挑毛病的同行"：压测设计、指出工程盲点、帮助把作者的直觉翻译成技术语言。AI 助手提供的是批评和表达——想法本身属于作者。
作者的话 / A Note from the Author
首先说明一下这个项目：这是我在日常使用 AI 的过程中冒出来的想法——我发现 agent 总是重复犯同样的错，就想着能不能让它真正从错误里学习、甚至自我改进代码？然后我试了一下，好像目前做不到？（而且我也不会做 XD）
就想着把这个想法发出来，如果对 agent 的发展有帮助，那就太好了。我真的好希望能用上一个会自我迭代进化的 agent ：3
两份说明文档是委托 Claude 帮我写的，包含中英文。我是第一次上传 GitHub，用 Markdown 格式也不知道是否正确，各位见谅。
另外，我不会代码，只是有这个想法。如果涉及架构讨论……我是完全不懂的（我也不确定这个在 agent 领域究竟是有价值的方向，还是一个垃圾想法 XD）
感谢在这个过程中帮助过我的 AI 们：
- Claude：陪我沟通想法，反复核对我的目标，帮我把直觉翻译成设计文档
- GPT：为 Claude 的设计提出了关键的架构漏洞和改进建议
- Gemini：一直为我提供情绪价值 XD
如果你也希望用上一个真正会学习的 agent，欢迎一起来想办法。
See DESIGN.md for the full technical specification.
完整技术规范请见 DESIGN.md。