💡 왼쪽 원문을 읽으면서 오른쪽에 따라 써보세요. Tab 키로 힌트를 받을 수 있습니다.

원문 렌더가 준비되기 전까지 텍스트 가이드로 표시합니다.

Introduction — We Read So Much, Yet Nothing Sticks

The HN front page, GeekNews, half a dozen newsletters, the article-sharing channel in the company Slack. We are the most-reading generation in history — yet ask most of us to explain an article we read a week ago, and we go blank. The confession post "I have 4,000 notes in Obsidian and cannot find anything" became a hit on Hacker News, and posts combining local LLMs with note systems keep climbing GeekNews — both are expressions of the same thirst.

There is another 2026-specific backdrop. With AI coding agents now ubiquitous, writing documents that **feed context to agents** — CLAUDE.md, AGENTS.md — has become standard practice. A well-organized personal note vault is no longer read only by humans; it becomes **the context source for your AI**. We have entered the era where your notes are your personal RAG (Retrieval-Augmented Generation) database.

This article distills the essentials of the classic PKM methodologies, clarifies what AI has changed, and then lays out a practical capture-process-connect-output system along with a 4-week build plan.

A Quick Tour of the PKM Classics — Zettelkasten and PARA

You do not need to read a whole methodology book. The essentials:

Zettelkasten — Connections Create Knowledge

The system with which sociologist Niklas Luhmann wrote 70 books from 90,000 note cards over a lifetime. Three core principles:

1. **Atomicity**: one idea per note. Only then do connections become possible.

2. **In your own words**: rewrite in your own sentences, never copy quotes. That act of translation is understanding itself.

3. **Connection first**: every time you create a note, ask "which existing note does this connect to?" Knowledge accumulates not inside notes but **between notes.**

PARA — Classify by Actionability

Tiago Forte PARA is a folder-structure theory. You keep only four buckets, organized by **actionability** rather than topic:

P - Projects : active work with deadlines (e.g., June blog redesign)

A - Areas : ongoing responsibilities with no end date (e.g., health, team ops)

R - Resources : topics of interest for someday (e.g., Rust, cooking recipes)

A - Archives : finished or inactive things

The sorting question is singular: "What does this note move right now?" If it moves a project, P. A responsibility, A. Mere interest, R. Nothing, Archives.

The Two Methodologies Compared

| Aspect | Zettelkasten | PARA |

| --- | --- | --- |

| Focus | Connecting and fermenting ideas | Execution and project momentum |

| Unit | Atomic permanent notes | Folders and documents |

| Strength | Writing, research, long-term thinking | Getting things done, retrieval speed |

| Weakness | High upfront cost, over-engineering risk | No deep connective structure |

| Best for | Writers and researchers | Practitioners juggling projects |

The punchline: **they are not competitors.** Manage working materials with PARA, ferment your thinking with Zettelkasten principles — mixing them is the realistic answer.

What AI Changed — Redefining the Role of Notes

From Search to Conversation

Retrieval used to mean keyword search and folder browsing. Now you embed your entire vault and **ask questions in natural language.** "What were the downsides of the saga pattern in that distributed transactions article I read last year?" is now a working query.

Automated Summaries — and Their Trap

In an era when AI summarizes your reading queue, the value of "copying out highlights" has dropped. But there is an important twist: **reading a summary and understanding are different things.** AI summaries are excellent as a capture-stage filter, but if you skip the processing stage of rewriting in your own words, no knowledge accumulates. Luhmann principle number two survives the AI era intact.

Notes Become RAG

The biggest shift. Notes have expanded from a human memory aid into **a context source for AI**:

[Before] me -> write notes -> future me searches and reads

[2026] me -> write notes -+-> future me searches / asks questions

+-> AI agents use them as context

(writing assistance, reference docs for

coding agents, automated weekly review

summaries, suggested note connections)

This adds one more criterion for a good note: **is it machine-readable?** A clear title, consistent metadata, a self-contained unit — which, conveniently, matches the Zettelkasten atomicity principle exactly.

The Tool Landscape — Local-First and Privacy

| --- | --- | --- | --- | --- |

The central 2026 debate is **privacy**. Connecting your notes to AI can mean sending your entire inner life somewhere. Three options:

1. **Cloud LLM + exclude sensitive notes**: convenient, but requires classification discipline.

2. **Local LLM (Ollama, llama.cpp and friends)**: you can safely connect even diaries and performance memos. As of 2026, local models are genuinely sufficient for summarization, tagging, and question answering.

3. **Hybrid**: embeddings and retrieval locally, long-form generation in the cloud.

My recommendation: **storage must be local markdown** (the notes outlive any tool), and start AI integration in hybrid mode.

Designing the Practical System — Capture, Process, Connect, Output

First, the whole pipeline in one picture:

+----------+ +----------+ +----------+ +----------+

| Capture | --> | Process | --> | Connect | --> | Output |

+----------+ +----------+ +----------+ +----------+

RSS/HN curation highlights to links/tags weekly review

read-later app permanent notes MOC building writing/talks

daily notes own-words pass AI suggestions asking the AI

Stage 1: Capture — the Reading Inbox

Principle: **separate discovery time from reading time.** Reading at the moment of discovery shatters your day.

- Designate one read-later tool (or a plain bookmarks folder) as **the only inbox.** Two scattered inboxes means both die.

- An HN/GeekNews curation routine:

HN/GeekNews routine (15 min/day, once in the morning)

1. Skim the front page, but never open an article (5 min)

2. Pick candidates by title + comment count, save to inbox only (5 min)

3. From yesterday inbox, choose only 1-2 things to read today (5 min)

Leave the rest. Items older than a week may auto-delete, and that is fine.

- Key rule: **the inbox is a candidate list, not an obligation.** The compulsion to read everything is what kills the system.

Stage 2: Process — Highlights Into Permanent Notes

What you highlighted while reading is not yet knowledge. Conversion rules:

Permanent note conversion rules (1-3 min per highlight)

1. Without looking at the highlight, write one paragraph in your own words

(if you cannot, you did not understand it -> reread or discard)

2. Title must be a complete claim sentence

Bad: "About caching"

Good: "Event-driven cache invalidation is safer than TTL"

3. Keep the source link

4. End by asking: "Which existing note does this connect to?"

Not every highlight deserves a permanent note. A realistic ratio is 1-2 out of 10. Let the rest flow away.

Stage 3: Connect — Link and Tag Strategy

- **Few tags, many links.** Keep tags as a small set for state management (e.g., seed, growing, evergreen). Use links and MOCs for topical organization, not tags.

- **MOC (Map of Content)**: once a topic exceeds about 10 notes, create a table-of-contents note that gathers them. More flexible than folders — one note can belong to several MOCs.

- AI leverage point: when you write a new note, having embedding search suggest "the 5 semantically closest existing notes" is a powerful workflow. AI discovers connection candidates; you make the judgment.

Stage 4: Output — Without It, the System Dies

- **Weekly review (a fixed 30 minutes)**: skim the notes created this week, empty the inbox, pick one note to develop next week.

- **Output through writing**: when 3-5 notes link together, you have the raw material for a blog post. This very article was assembled that way.

- The minimal unit of output need not be grand. Sharing a one-paragraph TIL in the team Slack is excellent output.

The Daily Note — The Heartbeat of the System

The engine that makes the capture-process-connect-output loop spin every day is the daily note. It starts from a single template auto-generated each morning:

2026-06-12 (Fri)

Today focus

- (one line, written last evening or this morning)

Work log

- 09:40 Reviewed the payment migration PR — liked the adapter boundary design

- 11:20 Saw a local-embeddings post on HN, sent to inbox

- 15:00 Incident retro meeting — the alert threshold discussion went in circles. Why?

TIL

- (3 lines: learned / struggled / note to tomorrow me)

Sent to inbox

- 2 links, 1 idea that surfaced during a meeting

For tomorrow

- Develop the alert-threshold thought into a permanent note

Operating principles:

- The daily note is a **throwaway note.** Do not try to write it well. Only the sentences that survive here get promoted to permanent notes.

- Timestamping the work log gives you your time-usage pattern for free at the weekly review.

- Stray thoughts during meetings also go into the daily note first. Getting them out of your head is itself what protects focus.

A Decision Tree for When Filing Stalls

A tree for deciding where new information goes within 3 seconds:

New information arrives

+-- Is it used by a project in progress?

| Yes -> straight into that project note in the Projects folder

+-- No -> Is it reading material?

| Yes -> into the inbox (judge later)

+-- No -> Is it my own thought/idea?

| Yes -> record it in the daily note for now

+-- No -> Is it reference material for someday?

Yes -> save just the link in Resources

No -> discard (discarding is also a function of the system)

This tree exists for exactly one reason: **to prevent filing decisions from breaking your flow.** If you deliberate longer than 3 seconds about where something goes, drop it in the daily note and move on.

AI Integration Recipes

Recipe 1 — Embed Your Notes Folder and Ask It Questions

Concept: split notes into chunks, store them as embedding vectors, embed the question too, retrieve the nearest notes, and pass them to an LLM as context. A minimal working example:

pip install chromadb ollama

Setup: ollama pull nomic-embed-text && ollama pull llama3.2

client = chromadb.PersistentClient(path="./note_index")

col = client.get_or_create_collection("notes")

1) Index the vault (paragraph-level chunks)

for path in glob.glob("vault/**/*.md", recursive=True):

with open(path, encoding="utf-8") as f:

text = f.read()

chunks = [c.strip() for c in text.split("\n\n") if len(c.strip()) > 80]

for i, chunk in enumerate(chunks):

emb = ollama.embeddings(model="nomic-embed-text", prompt=chunk)

col.upsert(

ids=[f"{path}:{i}"],

embeddings=[emb["embedding"]],

documents=[chunk],

metadatas=[{"source": path}],

)

2) Ask in natural language

question = "What did I write about the downsides of the saga pattern?"

q_emb = ollama.embeddings(model="nomic-embed-text", prompt=question)

hits = col.query(query_embeddings=[q_emb["embedding"]], n_results=5)

context = "\n---\n".join(hits["documents"][0])

sources = [m["source"] for m in hits["metadatas"][0]]

answer = ollama.chat(

model="llama3.2",

messages=[

{"role": "system",

"content": "The following are excerpts from the user personal notes. "

"Answer only from the notes; if it is not in the notes, say so."},

{"role": "user", "content": f"Notes:\n{context}\n\nQuestion: {question}"},

)

print(answer["message"]["content"])

print("Sources:", sorted(set(sources)))

In under 50 lines of code, "asking your notes" works. Everything runs locally, so indexing your diary is safe. Obsidian users can find community plugins that do the same job, but implementing the principle once yourself reveals the limits and possibilities of whatever tool you end up using.

Recipe 2 — A Summarization Prompt Template

A template for first-pass filtering of long articles at the capture stage:

Analyze the following article.

1. Core claim (3 sentences max)

2. The strongest and the weakest piece of evidence, one each

3. One counterargument the author does not address

4. Connection to my existing interests: which of

[distributed systems, developer productivity, LLM evals]

does this relate to, and why

5. Verdict: worth a deep read (high/medium/low) + one-line reason

Article: (paste body)

Item 4 is the point. Asking the summary to **judge connections to your interests** turns it from mere compression into a curation tool.

Recipe 3 — Automated Weekly Review Draft

Below are the notes created this week, with titles and bodies.

1. Is there a theme running through this week notes? What is it?

2. Suggest up to 3 pairs of notes that should be linked (with reasons)

3. Pick 1 seed note that could grow into an article

4. If any note from a month ago connects to this week themes, flag it

(paste note bodies or RAG context)

Caution: AI suggestions are drafts. Approving connections and assigning meaning is always a human job.

Preventing Note Debt — Guarding Against Perfectionism

The most common reason PKM fails is not laziness — it is **perfectionism**.

- **Do not try to finalize the folder structure first.** Structure is something you discover after 100 notes exist.

- **Do not turn every article into a note.** It is correct operation for 70% of the inbox to go unread and discarded.

- **Do not stall to unify note formats.** Ten badly formatted notes beat zero perfect ones.

- The diagnostic: **in the past month, did your notes improve at least one decision or deliverable?** If not, you do not have a system — you have a collecting hobby.

Three Developer-Specific Note Types

Code Snippet Notes

Title: Safely deleting duplicate rows in PostgreSQL

Tags: snippet, postgres

Context: used during the 2026-05 settlement table duplicate-load incident

Caution: CTID can change after VACUUM — trust only within a transaction

(code block)

Verified: tested on 2M rows in staging, 4.2s

The point is recording not just the code but **the context and the caveats.** Six months from now you will have forgotten all of it.

TIL (Today I Learned)

Three lines at the end of the day: "what I learned / what I struggled with / note to tomorrow me." The lowest-friction format, hence the most durable — and it becomes the source data for year-end retrospectives and resume updates.

Postmortem Notes

A personal-edition postmortem written after incidents and mistakes. Separate from the company document, it records "where exactly my judgment went wrong." The accumulated collection of these notes is what seniority actually consists of.

The 4-Week Build Plan

| --- | --- | --- | --- |

Verdict after 4 weeks: if the weekly review ran at least twice and at least one output exists, the system works. If not, do not switch tools — **halve the size of the routine** instead.

Anti-Pattern Collection

| Anti-pattern | Symptom | Prescription |

| --- | --- | --- |

| Tool hopping | Migrating note apps every quarter | Lock to markdown, pledge no migration for a year |

| Collector sickness | 3,000 inbox items, 0 permanent notes | Stop capturing, process only for a week |

| Perfect-structure compulsion | Endless folder redesigns | Treat structure as a discovered artifact |

| Highlight copier | All quotes, no own sentences | Rewrite-without-looking rule |

| Full AI delegation | Summaries pile up, understanding does not | Keep the processing stage manual |

| A garden with no harvest | Notes grow, nothing gets written | Put a monthly output slot on the calendar |

Pitfalls and Critical Perspectives

- **PKM itself can become productivity theater.** Tending the note system delivers the satisfaction of real work and can be the most sophisticated form of procrastination. Audit yourself by outputs.

- **"If AI remembers everything, why take notes?"** — a fair objection. Pure fact retrieval will increasingly be replaced by AI search. But the essential value of notes is not storage; it is **the clarification of thought that the act of writing forces.** That cannot be outsourced.

- **Embedding search is not a silver bullet.** Semantic search finds meaning across different phrasings, but exact keyword and code search is still better served by grep-like tools. They are complements.

- **Notes-as-RAG has a quality precondition.** Index garbage notes and you get garbage context — garbage in, garbage out. AI integration is an amplifier for a working processing stage, not a savior for a broken system.

Final Checklist

System

- [ ] Is there exactly one inbox

- [ ] Is the storage format local markdown (or something migratable)

- [ ] Are discovery time and reading time separated

- [ ] Is the weekly review fixed on the calendar

Note Quality

- [ ] Are permanent note titles complete claim sentences

- [ ] Are notes written in your own words, not quotes

- [ ] Does every note have at least one link

- [ ] Are sources preserved

AI Integration

- [ ] Is there a policy for sensitive notes (local-only / excluded)

- [ ] Is the embedding index refreshed periodically

- [ ] Is there a human-approval step for AI suggestions (links, summaries)

- [ ] Did you keep the processing (own-words) stage un-delegated

Health

- [ ] At least one output (post, decision, share) came from notes last month

- [ ] You have not switched tools within a year

- [ ] No compulsion to read the entire inbox

Closing

The flood of information will not stop, and AI will only make it flow faster. What makes the difference is not the ability to read more, but the loop of digesting what you read into your own words, connecting it, and producing output. Tools merely assist; AI is merely an amplifier. Today's task is simple: pick one inbox, and write one note in your own sentences. The second brain is built exactly that way — one card at a time.

References

- Zettelkasten Introduction — https://zettelkasten.de/introduction/

- Tiago Forte, The PARA Method — https://fortelabs.com/blog/para/

- Obsidian official site — https://obsidian.md/

- Logseq official site — https://logseq.com/

- Andy Matuschak, Evergreen notes — https://notes.andymatuschak.org/Evergreen_notes

- Ollama official docs — https://ollama.com/

- ChromaDB official docs — https://docs.trychroma.com/

- Sönke Ahrens, How to Take Smart Notes — https://takesmartnotes.com/

- Hacker News: note-taking discussion search — https://hn.algolia.com/?q=note-taking

- GeekNews (PKM and productivity discussions) — https://news.hada.io/