{"id":"2065475063928000681","url":"https://x.com/Av1dlive/status/2065475063928000681","text":"","author":{"name":"Avid","username":"Av1dlive","avatarUrl":"https://pbs.twimg.com/profile_images/2010392290590912512/x7vtAe5W_200x200.jpg"},"createdAt":"Fri Jun 12 16:43:24 +0000 2026","engagement":{"replies":27,"retweets":27,"likes":250,"views":344112},"article":{"title":"I built Obsidian from Scratch in 4 steps with 13+ Kimi Agents.","previewText":"I Built My Own Obsidian. Then put Kimi Work\n\nWhich Turned It Into a Second Brain.\nHere is what I Found\n \nTLDR; If you don't wanna read 3000 words of article simply give this to your agent","coverImageUrl":"https://pbs.twimg.com/media/HKoKmD1bEAE9eJq.jpg","content":"I Built My Own Obsidian. Then put Kimi Work\n\nWhich Turned It Into a Second Brain.\n\nHere is what I Found\n\n> TLDR; If you don't wanna read 3000 words of article simply give this to your agent https://github.com/codejunkie99/familiar-second-brain\n\n## Introduction\n\nThe problem is that most builders who try to build a second brain end up with a second filing cabinet.\n\nSome set up Obsidian because the graph view looks incredible in YouTube demos.\n\nSome try to wire Claude Code into their vault and never get past the initial setup friction.\n\nOthers jump straight into RAG pipelines and vector databases without understanding that the real problem is capture, not retrieval.\n\nThe result is always the same: a beautiful vault with 12 notes, a broken cron job, and a plugin graveyard.\n\nIf your goal is a knowledge system that actually compounds, you do not need RAG. You do not need a vector database. You do not need a cloud subscription.\n\nYou need to learn how to:\n\n- Structure a vault so any agent can navigate it without extra instructions\n\n- Write skill files that run repeatable workflows on a schedule\n\n- Use KIMI Work's agent to pull research directly into your vault while you sleep\n\n- Connect your vault as an MCP server so every agent you use reads from the same brain\n\n- Build a daily desktop page that surfaces what matters every morning\n\nThis article is the full build. Every skill file is included. Every cron expression is exact. Every architectural decision has a reason behind it.\n\nMy vault currently has 347 notes. A 50-file inbox batch processes in under 4 minutes with KIMI Work's Agent Swarm. The system has been running for 11 weeks.\n\nNow let's build it.\n\n## Before the Theory, Here Is What It Actually Does\n\nA raw note dropped into inbox/:\n\nWhat KIMI Work's agent writes back into notes/wiki/:\n\nThe agent created a structured wiki page, linked it to your existing knowledge graph, and surfaced a contradiction you had not noticed. You dropped a messy note. You got structured thinking back.\n\nThat is the loop. Everything below is how it works.\n\n## Why I Built Familiar\n\nThe Claude Code + Obsidian setup went viral in early April 2026. Over 50,000 likes in under a week. The idea was simple: point an AI agent at a markdown vault and passive note storage becomes active infrastructure.\n\nBut it has a ceiling.\n\n- Claude Code requires you to be present. You invoke it, it runs, you get output. It does not run at 7am while you sleep. \n\n- It does not pull research from the web on a schedule. It does not coordinate 300 parallel sub-agents to batch-process your document backlog.\n\nI wanted one vault that any agent could read from, write to, and maintain on a schedule. Local-first. Heavy lifting automated. I called it Familiar.\n\n![](https://pbs.twimg.com/media/HKoGG8EbIAAEn6q.jpg)\n\n## The Vault Architecture\n\nEverything lives as plain .md files on disk. No proprietary database. No cloud sync. \n\nMarkdown is the one format every AI agent reads and writes without friction\n\nThe resources/ folder never gets modified. Source material only gets ingested. \n\nThis prevents the failure mode that rots most AI knowledge bases: \n\n1. wiki content citing wiki content, \n\n1. the original source drifting further with every generation, \n\n1. the vault confidently believing things it fabricated in week two.\n\nThe AGENTS.md is the most important file in the repo. \n\nEvery agent reads it first. Claude Code, KIMI Work's agent, Cursor, any tool that respects the convention. One file, portable across the entire stack:\n\nRule 7 is the one that matters most. Without it, the agent fills knowledge gaps with plausible-sounding information that looks exactly like your real thinking. Your vault starts generating confident fiction\n\n## Building It: The Decisions That Actually Matter\n\n## Step 1: Version control before anything else\n\nBefore mounting the vault in any agent, run this:\n\n- I skipped this early on. An inbox processing task had an ambiguous instruction: \"merge related concepts where appropriate.\" \n\n- The agent decided my separate notes on [[Reasoning Models]] and [[Thinking Models]] were the same concept and combined them, discarding \n\n- 800 words of nuanced distinction written over three weeks. Two-line git revert. Lesson learned.\n\n## Step 2: Tiered memory, or why flat vaults become noise\n\nA flat vault breaks within weeks. The agent accumulates assumptions from projects that closed two months ago and synthesises from context that no longer reflects your actual priorities\n\nThe agent reads the right tier for the right kind of question. Stale warm-tier project notes never contaminate cold-tier synthesis.\n\n## Step 3: The YAML frontmatter convention\n\nEvery wiki page carries metadata the agent writes and maintains:\n\nhe last_updated_by field is how you audit the vault during weekly lint passes:\n\n1. Filter for last_updated_by: agent with confidence: low\n\n1. Read each page manually\n\n1. Correct, upgrade confidence, or delete\n\nTwenty minutes a week keeps the system clean.\n\n## Step 4: The inbox-first capture loop\n\nThe biggest friction in most note-taking setups is the gap between \"I found something interesting\" and \"it is in my system, linked and usable.\" Familiar's inbox pattern collapses this:\n\n1. Anything unprocessed goes into /inbox/ as a raw markdown file (web capture, voice memo, paste, or PDF drop)\n\n1. The agent reads the inbox on a scheduled trigger\n\n1. For each file: classify, assign to correct vault location, extract key concepts, generate [[backlinks]] to existing wiki pages, archive the original to /inbox/archive/\n\n1. Update index.md with the new additions\n\nSource files stay immutable. The agent creates a corresponding wiki page, but the origin is always traceable.\n\n## Inside the Familiar Build: Graph, Quick Capture, and Every Feature\n\nThe vault scaffold is the foundation. The features that make it feel like a real second brain rather than a folder of markdown files are in the layer built on top.\n\n## The Graph View\n\nFamiliar inherits Obsidian's [[wikilinks]] convention directly. Every backlink the agent creates becomes a bidirectional edge. Over time the graph reflects the actual shape of your thinking: which concepts cluster, which are heavily linked, where orphaned notes have never connected to anything.\n\nThe difference from vanilla Obsidian: the links are created by the agent, not by you manually. You write a raw note, the agent processes it, the graph grows. After 11 weeks and 347 notes, the clusters that have formed are not the ones I would have predicted when I started. That is the point.\n\nHow to read the graph:\n\n- Dense clusters are topics you have been thinking about consistently\n\n- Heavily linked single nodes are anchor concepts everything else references\n\n- Orphaned nodes are the most useful signal: thinking you did that never got integrated\n\n- Nodes with confidence: low are agent-generated pages that still need a human pass\n\n## Cmd+Shift+K: Instant Capture from Anywhere on Your Desktop\n\nThis is the feature that changed how I use the whole system day to day.\n\nKIMI Work's agent has a global hotkey bar. Hit Cmd+Shift+K from anywhere on your Mac and a command bar drops down. You do not need to switch windows. You do not need to open Familiar. You do not need to touch the terminal.\n\nYou type a thought, a question, or a half-formed idea, then tell it where to put it:\n\n1. KIMI Work's agent takes the input, formats it as a structured markdown note with a timestamp and source tag, and drops it directly into ~/familiar/inbox/. \n\n1. The next time the inbox processing cron fires at 8am, it gets classified, linked to your existing wiki pages, and archived like any other capture.\n\n1. The friction between having a thought and having that thought in your system drops to about four seconds.\n\nYou can also use the bar to query the vault on the fly:\n\nKIMI Work's agent reads your vault via the local file mount, answers from your actual notes, and you can tell it to append the answer as a new note or open the source file directly.\n\nPatterns I use with the quick bar every day:\n\n- save to inbox: thought captured mid-meeting, no context switching\n\n- add to crm/[person]:  something to remember about someone, added instantly\n\n- open project/[name] : jump directly to a project brief\n\n- flag conflict in wiki/[topic]: manually add a [CONFLICT] block without opening the file\n\nThe bar is modal and keyboard-native. For anyone who lives in the terminal, it feels exactly like the right abstraction.\n\n## Feature: Graph Health Skill\n\nThe graph health skill is an open issue actively in development. It runs on a weekly schedule and writes a full audit report back into the vault:\n\nRun this alongside the weekly lint pass. The health report tells you exactly where the vault is drifting and what needs a human decision before the next processing run. This is the feature that turns a static graph view into an active maintenance loop.\n\n## Feature: Voice Memo Ingestion\n\nVoice memo ingestion via a local Whisper instance is another open issue in active development. The intended flow:\n\n1. Drop a .m4a or .mp3 into /inbox/voice/\n\n1. KIMI Work's agent transcribes it via local Whisper\n\n1. Transcription is cleaned and saved as a standard markdown file in /inbox/\n\n1. The inbox processing skill picks it up on the next 0 8 * * * run\n\nThis closes the last manual capture gap. Right now the only way to get a thought into the vault while away from a keyboard is to voice-memo it and paste it manually later. The Whisper skill removes that step entirely. Combined with Cmd+Shift+K, every capture surface is covered.\n\n## Feature: WebBridge Reading List Capture\n\nCurrently in development as an open issue. Instead of manually maintaining a reading list file, WebBridge monitors a bookmarks folder or a shared URL list, auto-crawls new additions on the morning cron, and drops the full extracted text into /inbox/. The reading list becomes a queue. The agent drains it every morning.\n\n## Feature: MCP Server Exposure\n\n- Also an open issue. Once shipped, the full Familiar vault becomes an MCP server. Any MCP-compatible agent (Claude Code, Cursor, Windsurf, Hermes) can query it directly without separate setup. \n\n- You configure once. Every agent reads from the same brain. This is the feature that makes Familiar infrastructure rather than a personal tool.\n\n## Feature: Wiki Template Library\n\nA set of pre-built YAML frontmatter templates for the most common note types:\n\n- Person template : name, org, relationship, last contact, linked projects\n\n- Concept template : definition, related concepts, source, confidence\n\n- Project template : status, owner, deadline, linked people, linked concepts\n\n- Paper template : title, authors, published date, key claims, linked wiki pages\n\nEvery new page the agent creates will start from the right template instead of a blank file. Consistent structure means consistent graph quality from day one.\n\n## What Already Ships in the Repo Today\n\n- Vault scaffold with AGENTS.md template and index.md navigation map\n\n- Tiered memory structure (journal / projects / wiki / crm)\n\n- YAML frontmatter convention (confidence, last_updated_by, linked_sources)\n\n- Immutable resources folder with source-integrity rules\n\n- All three cron skill files (morning capture, inbox processing, weekly review)\n\n- Desktop page skill that generates a daily HTML briefing\n\n- Git-compatible flat file structure, no binary formats\n\n## KIMI Work Is Not a Chatbot. It Is a Desktop Agent.\n\nFor a while I was self-hosting Kimi K2.6 on my Mac mini. Pulled the weights via Ollama, ran it entirely offline, pointed my vault at it. \n\nZero cloud. Zero API costs. It worked well enough for querying notes but it was still me manually invoking it every single time I wanted something processed.\n\nKIMI Work changed that. It is the desktop application Moonshot built on top of K2.6, launched in June 2026. \n\n- Same model I was already running. But now it mounts my local folders, drives my browser, executes shell scripts in the background, and runs on a cron schedule without me opening a terminal. \n\n- The difference between self-hosting the model and running KIMI Work is the difference between having a tool and having a system.\n\nThe numbers that matter for this build:\n\n- 256K token context window : entire project folders fit in working memory\n\n- Agent Swarm : up to 300 parallel sub-agents, 4,000 coordinated steps, 4.5x faster than single-agent on large parallel tasks\n\n- Open weights under Modified MIT : still self-hostable via Ollama, vLLM, or SGLang if you want full offline\n\n- Native multimodal input : text, image, and video\n\nThe five capabilities\n\n1. Local File Agent\n\n1.  WebBridge\n\n1. Cron Scheduler \n\n1.  Agent Swarm \n\n1. Office Output\n\n## Skills: How KIMI Work's Agent Actually Runs Your Vault\n\nKIMI Work uses a skills system. A skill is a markdown file that defines a workflow. The cron job is the schedule. The skill file is the playbook. You edit the skill. The cron stays the same.\n\nThree skills run Familiar today.\n\nSkill 1: Morning capture (0 7 * * *)\n\nSkill 2: Inbox processing (0 8 * * *)\n\nSkill 3: Weekly review (0 18 * * 5)\n\nCron reference:\n\n## Connecting Everything via MCP\n\nKIMI Code CLI supports MCP. You can expose your Familiar vault as an MCP server, making it queryable from Claude Code, Cursor, any agent that speaks the protocol. One vault. Every agent.\n\nStep 1: Start the notes-mcp server\n\nStep 2: Register in KIMI Code CLI\n\nStep 3: Verify the connection\n\nStep 4: Query your vault from any agent session\n\nThe agent calls search_notes, reads the relevant wiki pages, and answers from your actual thinking, not its training data. The vault becomes infrastructure shared across every tool in your stack.\n\n## What Actually Compounds\n\nThis is what the 60-second demo videos never show.\n\nWeek 1\n\n- 40 notes\n\n- Agent creates structure from your first inbox batch\n\n- Useful, not magical\n\nWeek 4\n\n- 120 notes\n\n- Agent has real context about your thinking\n\n- When it processes a new paper, it cross-references four existing wiki pages, surfaces a contradiction from three weeks ago, flags it\n\n- You learn something about your own beliefs from the output\n\nMonth 3\n\n- 340 notes, hundreds of cross-links\n\n- Weekly briefings surface patterns you have not consciously noticed\n\n- Recurring open questions, themes you keep returning to\n\n- The orphan view shows exactly where you have been thinking without integrating\n\nThree rules that keep vaults compounding instead of rotting:\n\n1. Every wiki page cites at least one source from /resources/. No synthesis without a traceable origin.\n\n1. Weekly lint pass: filter last_updated_by: agent with confidence: low and review manually.\n\n1. Never let any agent write to /journal/. Your own timestamped thinking is the ground truth everything else builds from.\n\n## Get the Repo\n\nFamiliar is open source. The vault scaffold, AGENTS.md template, all skill files, the desktop page skill, and cron configuration ship ready to use.\n\n[github.com/codejunkie99/familiar-second-brain](https://github.com/codejunkie99/familiar-second-brain)\n\nThe full stack:\n\n- KIMI Work (local desktop agent): [kimi.com/products/kimi-work](https://www.kimi.com/products/kimi-work)\n\n- Kimi K2.6 open weights: [huggingface.co/moonshotai/Kimi-K2.6](https://huggingface.co/moonshotai/Kimi-K2.6)\n\n- notes-mcp vault server: [github.com/edvardlindelof/notes-mcp](https://github.com/edvardlindelof/notes-mcp)\n\n- Kimi Code CLI MCP docs: [moonshotai.github.io/kimi-cli](https://moonshotai.github.io/kimi-cli/en/customization/mcp.html)\n\nDrop questions in the replies. I read all of them.\n\nConclusion\n\nThis was made by the author's notes on the original prompt and edited by Kimi 2.6."},"adhxContext":{"savedByCount":1,"publicTags":[],"previewUrl":"https://adhx.com/Av1dlive/status/2065475063928000681"}}