Jinyoung Kim's AI4PKM System

Table of content

Jinyoung (Jin) Kim is Director of Naver Search US and Head of Data & Analytics at Naver, where he builds LLM-powered search engines. He previously led metric development and causal inference work at Snap and Microsoft across Bing, Outlook, and Cortana. He hosts the Data Intelligence podcast and wrote “Hello, Data Science,” a #1 bestseller in Korea’s technology category.

Kim created AI4PKM, a system where AI agents work alongside humans to maintain a personal knowledge base. Instead of treating AI as a chat interface, AI4PKM uses scheduled workflows and real-time task processing to keep your notes organized, enriched, and connected without manual effort.

Background

Kim’s career spans search infrastructure and data science:

At Microsoft, he worked on evaluation and measurement for web, email, and conversational search
At Snap, he led A/B testing and causal inference for recommendation systems
At Naver, he builds analytics platforms and an LLM-based search engine called Cue
Published over a dozen papers on information retrieval and data mining
Delivered tutorials on evaluation at SIGIR and serves on program committees for WSDM and WWW

His research background in search systems shows in AI4PKM’s architecture. The system treats knowledge management like a search engine problem: ingest, process, index, and retrieve.

The AI4PKM Methodology

AI4PKM operates on three principles that distinguish it from typical second brain systems:

Tool-agnostic design. The system assumes multiple tools will touch your knowledge base. Obsidian handles human editing. Cursor enables collaborative AI editing. Claude Code runs automated workflows. Each tool has a defined role instead of forcing everything through one interface.

Clear AI attribution. Every AI-generated piece of content includes labels showing which model created it and why. This lets you experiment with multiple AI tools without losing track of what came from where.

Scheduled maintenance loops. Instead of relying on chat interactions, the system runs automated workflows that process and organize content continuously.

The Four Workflows

AI4PKM bundles four automated workflows that run at different intervals:

Workflow	Frequency	Purpose
Daily Ingestion and Roundup (DIR)	Multiple times daily	Process photos, voice notes, web clips into structured notes
Weekly Roundup and Planning (WRP)	Weekly	Synthesize patterns and themes from daily roundups
Continuous Knowledge Upkeep (CKU)	Hourly	Fix broken links, update indexes, enrich content
On-demand Task Processing	Real-time	Handle web clips, voice commands, and #AI-tagged content

The Daily Ingestion workflow alone runs six steps: extract photo metadata, convert voice notes from Limitless.AI to markdown, enrich web clips with summaries and tags, generate daily roundups linking activities to topics, update relevant topic notes, and create social media posts from insights.

The system skips already-processed files using commit history and timestamps, so it runs efficiently even on large vaults.

Multi-Agent Architecture

The technical architecture moves beyond single-agent chat toward orchestrated agents. The system includes:

An Orchestrator that coordinates between specialized agents
A three-phase execution pipeline (KTG/KTP/KTE) for task generation, processing, and evaluation
Multi-agent routing that directs different tasks to appropriate agents
A CLI tool for triggering workflows manually or via cron

This mirrors how Jesse Vincent’s Superpowers separates concerns between brainstorming, planning, and execution agents. Kim takes it further by making the orchestration automated rather than human-triggered.

Obsidian Integration

The reference implementation uses Obsidian with pre-configured plugins:

Calendar and Dataview for structured queries
Templater for consistent note formatting
Tasks for action item tracking
Excalidraw for visual notes

The vault structure separates concerns: _Settings_/ holds prompts and workflows, Ingest/ handles incoming content, Topics/ and Projects/ organize permanent notes, and AI/ stores generated insights.

Kim also maintains claude-obsidian-skills, a collection of skills for Claude Code, Gemini CLI, and other assistants that support the AI4PKM workflows.

Key Takeaways

Principle	Implementation
Separate human and AI work	Different tools for different roles
Label AI content	Track which model created what
Automate maintenance	Scheduled workflows replace manual organizing
Process incrementally	Skip already-processed files
Build for multiple agents	Orchestrator coordinates specialized agents

Getting Started

Install the starter vault:

git clone https://github.com/jykim/AI4PKM

The repository includes documented prompts, workflow configurations, and the CLI tool for automated execution.

For lighter integration, the claude-obsidian-skills package works with existing vaults.

Links

Next: Tiago Forte’s Second Brain + AI