Graphify + Obsidian is INSANE: Build an AI Second Brain That Never Forgets — Transcript
Full transcript
- 0:00This is an AI second brain that never
- 0:02forgets. Point it at a folder and your
- 0:04whole project turns into a graph your
- 0:06assistant can query because right now
- 0:08your coding assistant has amnesia. Every
- 0:10new chat it has read nothing. Ask it how
- 0:13your login flow works and it opens every
- 0:15file from scratch to figure it out. Do
- 0:18that across 200 files on every single
- 0:20question and you pay a token tax each
- 0:22time. And stuffing your entire repo into
- 0:24a giant context window just makes the
- 0:26answers worse. On April 3rd, Andre
- 0:29Carpathy posted a fix and the dev world
- 0:31copied it in a single weekend. Stop
- 0:34making the model rebuild your knowledge
- 0:35every query. Give it a wiki it keeps up
- 0:38to date itself. 2 days later, a
- 0:40developer shipped exactly that as a tool
- 0:42called Graphify. From nothing, it blew
- 0:44past 90,000 GitHub stars in about 3
- 0:47months. Pair it with Obsidian, and that
- 0:49graph becomes plain Markdown you own.
- 0:51So, here's the full picture. How
- 0:53graphify builds the graph, why Obsidian
- 0:56is its home, and the four ways it falls
- 0:58apart if you get lazy. And this was not
- 1:00some quiet launch. It turned into a
- 1:02genuine wave. That claude code plus
- 1:05Obsidian idea pulled over 50,000 likes
- 1:07on X in a single week. One post
- 1:09declaring notetaking dead hit 54,000 on
- 1:12its own. Carpathy's little gist alone
- 1:15picked up 14,000 stars and,900 forks.
- 1:18First, the actual problem Graphify is
- 1:20solving. The usual fix for a big code
- 1:23base is what's called rag. Chop your
- 1:25files into chunks, turn them into
- 1:26vectors, and search for the closest
- 1:28matches when you ask a question. That
- 1:30works great for pros. For code, it's
- 1:32shaky because the link between a
- 1:34function and the things that call it
- 1:36isn't about similar words. It lives in
- 1:38the call graph. You end up paying to
- 1:40load the whole library just to find one
- 1:42paragraph. Graphify flips that around.
- 1:44Instead of searching text on every
- 1:46question, it does the expensive reading
- 1:48once up front and squeezes your project
- 1:50into an explicit graph of entities and
- 1:53the relationships between them. After
- 1:55that, answering a question means walking
- 1:57the graph instead of rereading the
- 1:58files. It's exactly how a senior
- 2:00engineer works. Build the mental map
- 2:02once, then just follow it. And it isn't
- 2:05only code. Graphify pulls in your SQL
- 2:07schemas, your shell scripts, your R
- 2:10notebooks, your architecture PDFs, even
- 2:12recorded videos, and drops them into one
- 2:15graph. Application code, database
- 2:17schema, and infrastructure finally
- 2:19living on the same map instead of in
- 2:21five different heads. Actually using it
- 2:23is almost boring in the best way. Inside
- 2:26cloud code or codeex or cursor, you type
- 2:29/graphify and point it at a directory.
- 2:32It grinds through the folder and drops a
- 2:34graphify out folder right next to your
- 2:36code with the entire graph saved to
- 2:38disk. It supports every big assistant,
- 2:40not just one. Here's the part one
- 2:42genuinely respect. Graphify does not
- 2:45treat every file the same way. It runs
- 2:47three separate passes over your stuff
- 2:49and each one has a completely different
- 2:51story about where your data ends up.
- 2:53Pass one is your source code. Graphy
- 2:55runs it through tree sitter, the same
- 2:57kind of parser a compiler uses across 40
- 2:59languages. It's deterministic. There's
- 3:02no AI model involved and there's no
- 3:04network call. Your code stays on your
- 3:06machine and every link it finds gets
- 3:08tagged as a hard fact. Pass two is for
- 3:11recordings. Point it at a folder with
- 3:13audio or video, say a design meeting you
- 3:15screen recorded, and it transcribes them
- 3:17locally with faster whisper. An hour of
- 3:20talking becomes text right on your
- 3:21laptop and folds into the graph. Nothing
- 3:24gets uploaded. Pass 3 is the honest one.
- 3:27Documents and images, a PDF spec, or a
- 3:30photo of a whiteboard can't be read by a
- 3:32grammar. So, those get sent to your AI
- 3:34provider to be understood. That's the
- 3:36only pass that leaves your machine, and
- 3:38it goes to your own API key, not some
- 3:40middleman. Want it fully private? Run
- 3:43code only mode and skip this pass. And
- 3:45it refuses to sound sure about things it
- 3:47isn't. Every connection in the graph
- 3:49carries a label. Extracted means it's
- 3:51right there in your code, so trust it.
- 3:53Inferred means the model reasoned it
- 3:55from context. So probably ambiguous
- 3:58means the model itself raised a hand. So
- 4:00check before you build on it. Once the
- 4:02graph exists, it organizes itself.
- 4:05Graphify runs community detection and
- 4:07your modules just fall out of the shape
- 4:09of the code. O in one cluster, billing
- 4:11in another, infrastructure off on its
- 4:14own with the bridge files between them
- 4:15lit up. You didn't draw that map. The
- 4:18relationships did. Now the payoff at
- 4:20query time. You ask, "How does a user
- 4:23log in?" Instead of loading 50 files,
- 4:25Graphify walks the graph and hands back
- 4:28a short path. Three hops, zero files
- 4:30opened. Your assistant reads a few node
- 4:32summaries instead of half the repo. And
- 4:34querying isn't just one trick. Graphify
- 4:37explain hands you a plain English tour
- 4:39of any entity. Graphy path finds the
- 4:42shortest chain between two things, say
- 4:43auster and the database pool it secretly
- 4:46leans on. and deep mode goes hunting for
- 4:48the fuzzier inferred links a quick pass
- 4:50would skip right over. Graphy puts a
- 4:53number on all this about 71 times fewer
- 4:55tokens per query on a mixed codebase.
- 4:58Now that's their own benchmark on their
- 5:00own data and nobody outside has
- 5:01reproduced it yet. So take the exact
- 5:03figure with a grain of salt, but the
- 5:05direction is obviously right. Walking a
- 5:08map is cheaper than rereading the whole
- 5:09library every time. And that one graph
- 5:12isn't trapped inside the tool. Graphy
- 5:14exports it every way you'd want. a
- 5:16clickable HTML graph rendered with viz.s
- 5:19s plain JSON for scripts, a markdown
- 5:22report of your busiest files, a Neo4j
- 5:24database, even an MCP server, so any
- 5:27agent can call your codebase graph like
- 5:29a built-in tool. It also stays current
- 5:32without you babysitting it. Every file
- 5:34gets a chat 256 hash, so the second run
- 5:37only rereads what actually changed. Add
- 5:39a git hook and the graph rebuilds itself
- 5:42on every commit. Your assistant sees a
- 5:44map that matches the exact branch you're
- 5:46standing on. So, where does Obsidian
- 5:48fit? Obsidian is a notes app built on
- 5:51plain markdown files that live on your
- 5:53disc, not on someone's server. Around 1
- 5:56and a.5 million people use it, and its
- 5:58plugins just crossed 100 million
- 5:59downloads. It's the natural home for a
- 6:01brain you actually want to own. Point
- 6:03Graphifies Obsidian export at a vault,
- 6:06and your graph turns into real notes.
- 6:08Every entity becomes a page. Every
- 6:10relationship becomes a Wiklink. And
- 6:12Obsidian's graph view shows the whole
- 6:14thing as a web. You can wander through
- 6:16your codebase as a mind map. You can
- 6:18open on a Sunday. Then you close the
- 6:20loop with a plugin like Claudian, which
- 6:22drops Claude code straight into the
- 6:23Obsidian sidebar. Your vault becomes the
- 6:25agents working directory. It can read
- 6:27your notes, write new ones, and run
- 6:29commands without leaving the app your
- 6:31knowledge already lives in. And Graphify
- 6:33isn't the only door in. You don't have
- 6:36to hand over full write access on day
- 6:38one. Smart connections just surfaces
- 6:40related notes using embeddings. Cudge is
- 6:43the self-hosted run it all locally
- 6:44option. Copilot for Obsidian gives you
- 6:47vault aare chat and is already past
- 6:49100,000 users. Start light then give it
- 6:52more rope. Underneath all of this sits
- 6:54Carpathy's actual design and it's
- 6:56refreshingly simple. Three layers. A raw
- 6:59folder of sources the AI leaves
- 7:01untouched. A wiki layer the AI does own
- 7:04where it writes pages and keeps the
- 7:05cross links honest. And a schema file
- 7:08that tells the model how your wiki is
- 7:09laid out. Three operations keep that
- 7:11wiki alive. Ingest pulls in new sources
- 7:14and updates every page they touch. Query
- 7:17searches the wiki and answers you with
- 7:18citations. And lint runs a health check
- 7:20for contradictions, dead pages, and
- 7:23stale claims. Remember that word lint.
- 7:25It comes back to bite people. And what
- 7:27gets me is how old this idea actually
- 7:30is. Carpathy traces it back to 1945 to
- 7:33Vanavar Bush's memes. The dream of link
- 7:35trails between your documents. The links
- 7:37mattered as much as the pages
- 7:39themselves. We just lacked a tool
- 7:40patient enough to maintain them. Now the
- 7:43model does that upkeep for free. Okay,
- 7:45I've talked this up for 5 minutes, so
- 7:47let me be the annoying person in the
- 7:49comments for a second. The viral demos
- 7:51leave out the part where this whole
- 7:52thing rots if you ignore it. There are
- 7:54four cracks worth knowing about. Crack
- 7:56one is size. A vault with 50 tidy notes
- 8:00feels like actual magic. A vault with
- 8:022,000 notes turns into a search problem
- 8:04all over again because the model spends
- 8:06its context just navigating your
- 8:08structure instead of answering your
- 8:10question. The graph softens that. It
- 8:12doesn't erase it. Crack 2 is confident
- 8:15fiction. When the assistant writes up
- 8:16your notes, it cheerfully fills the gaps
- 8:19with things you never said in your own
- 8:20voice. The fix is a blunt line in your
- 8:23instructions. Do not add anything I
- 8:25didn't write. Without that rule, your
- 8:27second brain starts inventing memories.
- 8:29Crack three is feedback loops. The AI
- 8:32writes a page, then weeks later cites
- 8:34its own page as a source, then builds on
- 8:36top of that. Let it run unattended and
- 8:38you get AI quoting AI where small
- 8:40mistakes slowly harden into facts. Your
- 8:43brain fills up with its own echo. Crack
- 8:454 is the big one, and it's why that word
- 8:47lint mattered. Carpathy built a cleanup
- 8:50pass into the design on purpose, and
- 8:52almost every tutorial skips it. 5
- 8:54minutes a week, read what the agent
- 8:56wrote, prune the stale pages, kill the
- 8:58contradictions. Skip that and the whole
- 9:01thing decays into noise. Do it and it
- 9:03compounds into something real. So, my
- 9:06honest take, a second brain that never
- 9:08forgets is genuinely here and it's the
- 9:10best version we've had. But that memory
- 9:12cuts both ways. It also keeps every bad
- 9:15note and lazy guess unless you clean up
- 9:17after it. Own your markdown, keep it on
- 9:19your disc, and treat that weekly lint
- 9:21pass as the rent. And if you do one
- 9:23thing this week, start almost
- 9:25embarrassingly small. One folder, one
- 9:27graph, one vault. Let it earn your trust
- 9:30on a project you already understand
- 9:32before you feed it your entire life. The
- 9:34people who win with this aren't the ones
- 9:35with the sickest demo. They're the ones
- 9:37whose vault still tells the truth 3
- 9:39months from now. That's the whole build.
- 9:42A brain that finally keeps up with you.
- 9:44See you in the next one.
About this transcript
This page contains the full transcript of Graphify + Obsidian is INSANE: Build an AI Second Brain That Never Forgets by Cloud Codes, generated from the public captions YouTube serves with the video. The transcript has 1,824 words across 265 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.
What you can do with it
Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to translate the transcript into 100+ languages.
Free YouTube transcript tool
YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, translation to 100+ languages, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.