Skip to content

Instantly share code, notes, and snippets.

@PerrinYong
Last active September 26, 2026 13:23
Show Gist options
  • Select an option

  • Save PerrinYong/c9c231cb1a96c22a7244c67d817f5997 to your computer and use it in GitHub Desktop.

Select an option

Save PerrinYong/c9c231cb1a96c22a7244c67d817f5997 to your computer and use it in GitHub Desktop.
See the Forest · 见林 — A pattern for building, evolving, and publicly exploring personal knowledge networks with LLM Agents / 使用 LLM Agent 构建、演化并公开浏览个人知识网络

See the Forest · 见林 — A Knowledge Evolution System

See the Forest · 见林

Language: 中文 · English

A pattern for building, evolving, and publicly exploring a personal knowledge network with LLM Agents.

Product Website · Explore the Live Knowledge Network · Builder on GitHub

Core Idea

In the age of LLMs, knowledge has become remarkably easy to obtain. So what value is left in maintaining personal notes and a personal blog over the long term?

I began taking notes as a student and have accumulated more than four thousand documents over the years. What is worth preserving is not only what each document says, but also how the documents connect, how they corroborate one another, and how they have changed my understanding.

None of this exists in any single, isolated note. It exists between the notes.

Karpathy's LLM Wiki was a major inspiration: let an Agent maintain a Wiki incrementally, so that high-value results become lasting artifacts that continue to compound over time.

See the Forest draws on that essential idea, but it does not aim to place another layer of summaries on top of the source material. I believe that the network formed by knowledge is itself part of the knowledge. The original notes, the relationships between them, and the understanding that emerges as those relationships evolve should form a whole.

The value of an Agent is not that it thinks for me. Its value is that, through greater capacity and efficiency, it makes forms of knowledge maintenance possible that were previously too difficult to sustain. I call the gradual emergence of new connections across time, sources, and fields "knowledge fermentation."

A model can give me answers, but it cannot build my understanding for me.

Architecture

1. Knowledge Foundation

The first step is to revisit and reasonably classify the notes accumulated over time, then build a directory tree that can be maintained for the long term, so that every note has a basic home. Existing Markdown notes and directories remain the knowledge foundation; adopting this approach does not require rebuilding everything from scratch.

But a directory can only determine where a note belongs. It cannot fully express the relationships between ideas.

2. Knowledge Evolution Layer

The knowledge evolution layer maintains semantic relationships, synthesized judgments, evidence boundaries, contradictions, open questions, and the evolution of understanding. It is not a separate Wiki detached from the original notes. Together with the knowledge foundation, it forms one network.

Maintenance here does not mean allowing an Agent to rewrite my words at will. An Agent can help discover relationships, compare new knowledge with what came before, and maintain the network, but the final judgment remains mine.

3. Public View

Once the knowledge network has taken shape, it can also have a public window. In the See the Forest blog, the knowledge network is not a decorative graph beside a conventional article list. It is the primary entrance. Every node represents real public content, and every connection comes from a real knowledge relationship.

The network is like a field of stars made of knowledge. The stars connect to one another, and the strength of those connections reflects the strength of their relationships. You can drag the network with a mouse, or select a node to enter a topic or article.

In use, the nodes move naturally and the connections remain clear. Even with a large amount of content on screen, browsing and dragging stay fluid. This document intentionally does not describe the underlying implementation.

Operation

I call the process through which an Agent helps knowledge continue to evolve "knowledge compilation":

  1. Locate: Find the existing knowledge most relevant to new material or a new question.
  2. Compare: Determine what the new material adds, supports, contradicts, constrains, or changes.
  3. Write back: Preserve the changes that have lasting value in the knowledge network.
  4. Validate: Check the sources, boundaries, and whether the proposed changes are acceptable.

Navigation and Evolution History

I use classification, bidirectional links, and Graph View to gradually connect previously scattered material. But more links are not necessarily better. What matters is why two ideas are related, and how new material changes what was already understood.

Optional: A Public Entry Point for Agents

In the age of LLM Agents, humans are no longer the only readers of the internet. This blog should therefore be easy for people to read, while also allowing Agents to discover, retrieve, and understand its content.

This does not mean adding a chat box to the homepage. Rather than having a model answer in my place, I care more about whether an Agent can read my original words, understand the relationships between ideas, and bring every citation back to the specific content it came from. People and Agents may enter through different paths, but they should ultimately see the same content and the same knowledge network. Private knowledge never enters the public blog.

Why This Works

LLM Agents have not changed the nature of knowledge. They have made forms of long-term maintenance practical by increasing the amount of work that can be done. People still ask the questions, judge what matters, and make the final decisions. Agents help discover relationships across large bodies of material, compare new and existing knowledge, and maintain the network.

The blog gives that network a window through which it can be seen: not only a single tree, but also the relationships between the trees and the forest they form together.

Note

This document intentionally stays at the level of ideas and visible effects. It does not describe the underlying implementation of the knowledge network's visual and interaction system. It presents only one pattern:

Let knowledge do more than remain stored: let it connect, continue to evolve, and become visible again in forms that work for both people and Agents.

Author and other projects: Perrin Yong

Product website: seetheforest.net

Live knowledge network: pystone.net

Inspired by: Karpathy — LLM Wiki

See the Forest · 见林 — A Knowledge Evolution System

See the Forest · 见林

语言: 中文 · English

一种使用 LLM Agent 构建、演化并公开浏览个人知识网络的模式。

产品官网 · 体验真实知识网络 · 构建者 GitHub

核心思想

LLM 时代,知识的获得已经变得极为容易。那么,长期维护个人笔记和个人博客,还有什么价值?

我从学生时代开始记录笔记,至今陆续积累了四千多篇文档。真正值得保留的,不只是这些文档各自写了什么,还有它们怎样连接起来,怎样互相印证,以及怎样改变了我的认识。

这些东西不在任何一篇孤立的笔记里,而存在于笔记之间。

Karpathy 的 LLM Wiki 给了我很大启发:让 Agent 持续增量地维护 Wiki,把高价值结果沉淀为可以长期存在、不断产生复利的制品。

See the Forest 借鉴了这个本质,但不希望只在原始资料之外增加另一层摘要。我认为,知识构成的网络,也是知识的一部分。原始笔记、知识之间的关系,以及这些关系长期演化形成的认识,应该构成一个整体。

Agent 的价值不是替我思考,而是从工作量和效率上,让过去很难长期执行的知识维护变得可行。我把知识在跨时间、跨材料和跨领域的互动中逐渐产生新联系的过程,称为“知识发酵”。

模型可以给我答案,但给不了我的认知。

架构

1. 知识底座

第一步,是先把过去积累的笔记重新梳理、合理分类,构建一棵能够长期维护的目录树,让每篇笔记先有一个基本归属。现有的 Markdown 笔记和目录仍然是知识底座,不需要为了采用新方案而推倒重建。

但目录只能决定一篇笔记放在哪里,不能完整表达知识之间的关系。

2. 知识演化层

知识演化层维护语义关系、综合判断、证据边界、矛盾、开放问题和认识演化。它不是另一套与原始笔记分离的 Wiki,而是与知识底座共同构成一张网络。

这里的维护,并不意味着让 Agent 任意改写我的原文。Agent 可以帮助发现关系、比较新旧知识和维护网络,但最终判断仍然由我完成。

3. 公开观察层

知识网络形成以后,它还可以拥有一个公开的观察窗口。在 See the Forest 博客中,知识网络不是传统文章列表旁边的一张装饰图,而是主要入口。图中的节点对应真实公开内容,连接来自真实的知识关系。

这个网络,就像由知识构成的一片星辰。星辰之间互相连接,也会根据关系的强弱呈现不同的连接强度。你可以用鼠标拖动,也可以点击一个节点进入主题或者文章。

从效果上看,节点移动自然,连接关系清晰,即使同时展示大量内容,浏览和拖动也依然流畅。具体如何实现,本文不展开。

操作

我把 Agent 协助知识持续演化的过程称为“知识编译”:

  1. 定位:找到与新材料或问题最相关的旧知识;
  2. 比较:判断它新增、支持、反驳、限制或改变了什么;
  3. 写回:把真正有长期价值的变化留在知识网络中;
  4. 校验:检查来源、边界和修改是否可以接受。

导航与演化记录

我借助分类、双向链接和 Graph View,把分散的内容逐渐连接起来。但链接不是越多越好,真正重要的是它们为什么相关,以及新材料怎样改变已有认识。

可选:面向 Agent 的公开入口

在 LLM Agent 时代,人类已经不是互联网唯一的读者。因此,这套博客不仅要方便人类阅读,也要方便 Agent 发现、读取和理解。

这不是在博客首页增加一个聊天框。相比让模型代替我回答,我更在意 Agent 能不能读到我的原文、理解知识之间的关系,并把引用带回具体内容。人和 Agent 可以从不同入口进入,但最终看到的,应该是同一份内容和同一张知识网络。私有知识不会进入公开博客。

为什么这种方式有效

LLM Agent 并没有改变知识的本质。它只是通过工作效率的提升,使过去很难长期执行的知识维护变得可行。人仍然负责提出问题、判断价值和作出最终决定,Agent 负责从大量内容中发现关系、比较新旧知识并维护网络。

博客则让这张网络获得了一个可以被看见的窗口:不仅看到一棵树,也看到树木之间的关系,以及它们共同构成的森林。

说明

本文有意停留在理念和可见效果层,不介绍知识网络视觉与交互系统的底层实现。它想表达的只有一个模式:

让知识不仅被保存,还能够彼此连接、持续演化,并以适合人和 Agent 的方式被重新看见。

作者与其他项目:Perrin Yong

产品官网:seetheforest.net

真实知识网络:pystone.net

灵感来源:Karpathy — LLM Wiki

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment