Agentic Zettelkasten Skill

A structured Zettelkasten protocol for Advanced Memory that produces high-quality atomic knowledge cards through deliberate retrieval, enrichment, and explicit link decisions.

This skill replaces ad-hoc note-taking. Every card written through this workflow is an atomic insight in your own words, with [[wikilinks]] embedded in sentences explaining the relationship — not a bare tag list.

Core Principles (Luhmann's Method)

1. Atomicity — One insight per card. If you have two ideas, write two cards.
2. Own words — Distill and rephrase. Copy-paste = no understanding (Feynman test).
3. Links in context — [[links]] must appear in sentences explaining the relationship. Good: "This contradicts [[jwt-migration]] — stateless tokens can't be revoked." Bad: "Related: [[jwt-migration]]"
4. Retrieve before write — Always search for existing cards first. Never create duplicates.
5. Preview before mutate — Show planned changes before writing anything.

Tools Available

Workflow

Step 1: Observe — What did we learn?

After completing a task, identify reusable insights. Skip if the content is:

• Temporary or session-specific (in-progress work, task state)
• Obvious or well-known (standard library usage, common patterns)
• Not reusable across future sessions

If nothing is worth remembering, stop here. This is a quality gate, not a formality.

Step 2: Draft — One atomic card per insight

For each insight, draft exactly one card:

• Title: Short noun phrase (≤60 chars, kebab-case slug). Descriptive, not clever.
• Body: One atomic insight in your own words. 3-8 sentences. Include:
  • What the insight IS
  • Why it matters (context)
  • When it applies (boundary conditions)
  • A concrete example or counterexample

If you need more than 8 sentences, you probably have two insights. Split them.

Step 3: Enrich — Proposed metadata

---
title: <Short Noun Phrase>
created: <YYYY-MM-DD>
category: <one category: developer|researcher|writer|knowledge-worker|etc.>
tags: [<3-5 comma-separated categories>]
keywords: [<3-8 salient terms>]
---

Step 4: Retrieve — Search before writing

Search for related existing cards using at least TWO queries:

adn_search(operation="rag", prompt="<topic keywords>", limit=5)
adn_search(operation="query", text="<exact term>", search_type="title")

Read the top candidates that look relevant:

adn_notes(operation="read", identifier="<slug>")

Limits to avoid runaway retrieval:

• Max candidate searches: 3
• Max cards read: 10

Step 5: Decide — create, update, or skip

Choose exactly one action per insight:

Rules:

• Links must be chosen by reading candidate cards, not from embedding scores.
• Updating is allowed only after preview (Step 6).
• If in doubt between create and update: create. It's safer to have two cards than to corrupt one.

Step 6: Preview — Show planned changes

Before writing, produce a preview:

Planned memory changes:
- CREATE: python-asyncio-gotcha (title: "Asyncio gather() doesn't guarantee order")
  → links: [[task-queue-pattern]] because tasks run concurrently, not sequentially
  → metadata: category=developer, tags=python,async,concurrency
- UPDATE: task-queue-pattern — add note about gather() ordering pitfall
  → append: "Note: asyncio.gather() starts all tasks concurrently; results
    may arrive out of order. Use asyncio.as_completed() for ordered results."

Step 7: Write

Create new card:

adn_notes(operation="write",
  title="Asyncio gather() doesn't guarantee order",
  content="...",
  folder="developer",
  tags="python,async,concurrency")

Update existing card:

adn_notes(operation="edit",
  identifier="task-queue-pattern",
  mode="append",
  content="...")

Rules for writing:

• Wikilinks go in body text, not just frontmatter
• Every link has a "because" clause
• Title must match slug (kebab-case)
• Content is markdown with at least one [[wikilink]]

Step 8: Verify — Readback check

After writing, read the card back:

adn_notes(operation="read", identifier="<slug>")

Confirm:

• Title and body match what was intended
• All [[wikilinks]] are syntactically valid
• Metadata frontmatter is complete (title, created, category, tags)
• Body is in your own words (not copy-pasted)

Anti-Patterns