Methodology

Based on Andrej Karpathy’s LLM Wiki — KnowFlow is an engineering attempt at that idea.

The core problem

Every mainstream answer to “how do I keep what I learn?” fails the same way:

Approach	Problem
Bookmarks / read-later	The “saved = learned” illusion (Collector’s Fallacy)
Note apps	Solve storing, not organizing or connecting
Plain RAG	Search from scratch every time — no memory, no accumulation

The core insight

Stop digging through the original documents every time. Compile the knowledge once, keep it ready, and reuse it.

An analogy:

• RAG is a temp worker — every question sends it scrambling through the filing cabinet, stapling scraps into an answer, remembering nothing for next time.
• An LLM Wiki is a full-time librarian — it spends its time extracting information, finding connections, building indexes, and maintaining a knowledge graph. When you ask, it retrieves.

Three layers

1. Raw layer — the material

Everything you feed in: articles, tweets, PDFs, WeChat posts, YouTube transcripts, notes.

Rule: the raw layer is immutable storage. Capture never rewrites it.

2. Wiki layer — the compiled knowledge

Structured, linked pages built from the raw layer:

• Entity pages — concrete things: people, companies, projects
• Concept pages — abstract ideas: methods, techniques, patterns
• Comparison pages — systematic A-vs-B comparisons
• Source pages — traceable records of where knowledge came from

The key property: pages link to each other. Mention a concept and it points at that concept’s page. Over time, the wiki becomes a living network — not a flat pile of snippets.

3. Schema layer — the rules

The “job manual” that tells any agent how to compile: what to extract, what to ignore, page formats, linking policy. In KnowFlow this is the page templates, the data-model reference, and the synthesis contract at docs/agents/synthesize.md — inspectable, versioned Markdown.

Why this beats plain RAG

RAG:
  question → search raw documents → staple an answer → forget everything

LLM Wiki:
  new content → compiled into wiki pages → linked → graph updated
  question → retrieve from compiled knowledge → answer with full context

1. It has memory — knowledge accumulates instead of restarting from zero
2. It has structure — entities, concepts, comparisons — not flat fragments
3. It’s explorable — the graph surfaces connections you didn’t search for
4. It’s verifiable — every page traces back to sources

Where KnowFlow fits

KnowFlow is an open-source implementation of this methodology:

• A CLI that keeps capture deterministic and the raw layer immutable
• Templates that define the schema layer
• Local tools for the graph, health checks, and optional vector search
• Synthesis deliberately delegated to any coding agent following the written contract, with human review as the gate — because that is where judgment lives

The core belief: the most interesting part is not the final tool — it is watching scattered information turn into structured knowledge, one page at a time.