note.md - your notes and research documentation now a local LLM Memory
by•
A local-first research workspace for Mac. Read papers, manage sources, take markdown notes, cite evidence, and turn literature into structured writing — instead of juggling Zotero, Obsidian, PDF readers and writing apps.
Replies
Huge congrats on the update! Keeping the vault as plain Markdown files while opening it up to AI is a game-changer.
Since it reads the folder structure directly, do you have any tips for how to organize notes to make it easiest for an agent like Claude to navigate?
note.md
@doganakbulut
I can't give you a real best practice, but what I learned so far:
The usage of wiki links is a game changer. Claude does not receive the graph that the user sees but It sees that there is a connection to another article and will most likely load that into the context as well. So keeping your notes connected where topics are overlapping is something I would definitely recommend.
the support-vs-contradiction scan is the sharp bit — embeddings sit 'x causes y' next to 'x doesn't', so retrieval finds candidates but stance needs an nli pass on top. on-device per claim is the real cost.
note.md
@harshchandgotia
That distinction is basically the whole thesis of note.md. A note is a means; the thing you're actually building is a connected body of understanding — and most tools optimize for capture (get the note down) while quietly neglecting the part where it becomes knowledge.
A few choices fell out of taking that seriously. Sources aren't an afterthought — a claim is linked to the evidence it rests on, because knowledge you can't trace back isn't really yours yet. The graph and links exist so structure is visible, not just storage. And the AI is deliberately a librarian, not a ghostwriter: it surfaces what you've read and the evidence for and against a claim, but it doesn't write your prose — because the deliberate thinking is the research, and outsourcing it hollows out the exact part that builds knowledge.
So the bar I hold the whole app to is: does this help understanding accumulate, or does it just help notes pile up? Capture is easy. Compounding is the hard, interesting part.
Congrats on the launch! Keeping the vault as an open, flat directory of plain Markdown files is a huge win for portability. I'm curious about the background file-watching mechanics. If a user modifies their .md files or directory structure externally via terminal or another editor like Obsidian, note.md seamlessly detects and re-indexes those changes on the fly, or is a manual re-sync required to keep the reference manager and source connections aligned?
note.md
@juno_dost
this is something note.md is specifically designed to handle, so external edits are a first-class case, not an afterthought.
Short answer: it's automatic. No manual re-sync needed.
note.md runs a live file-system watcher (built on macOS FSEvents) against your vault directory. When you edit a .md file in Obsidian, change something from the terminal, add or delete files, or restructure folders, the watcher picks that up and kicks off an incremental merge in the background — the sidebar tree, the reference manager, and the graph connections all realign on their own.
A few details worth being transparent about:
- It's incremental, not a full rescan. Each file is fingerprinted by modification time and size, so unchanged files are skipped and only what actually changed gets re-indexed. That keeps it fast even on large vaults, and the update carries through to exactly the articles that were touched — so wikilinks and source connections stay consistent without rebuilding everything.
- An open editor won't get clobbered. If you have a note open in note.md and it changes on disk underneath you, note.md does a three-way reconcile rather than blindly overwriting. If the changes don't conflict it fast-forwards silently; if they do, you get a conflict banner so you decide. You won't lose work to a background sync.
- The on-disk files are the source of truth. note.md treats your vault directory as canonical and mirrors it, which is exactly why editing in Obsidian or via terminal "just works" — there's no separate database you have to manually reconcile against.
So in practice: edit wherever you like, however you like, and note.md keeps the reference manager and source graph aligned on the fly. Manual re-sync exists as a fallback, but day to day you shouldn't need to reach for it.
Loomal
Love the local-first approach — keeping your notes as LLM memory is a genuinely clever solve for the 'AI doesn't know my work' problem. Does note.md support linking between docs to build a knowledge graph over time?
note.md
@dannyheng Yes, and if you wish to do so we support exporting to Neo4J format :)
Oh my gosh, I've struggled with citation managers for so long! And this is built right into Notes. It's native to the Mac, too, so it runs smoothly. Thank you!
Congrats. What has surprised you most from early feedback today?
AISA AI Skills Test
the vault-as-LLM-memory angle is really smart. most AI note tools try to be the AI themselves — this just makes your existing research available to whatever model you're already using. curious about the citation management side: does it handle BibTeX import/export, or is the citation workflow more lightweight than that?
note.md
@ozandag
You can export your sources from the citation management as .bib file. But import only works via pdf or link so far
The plain-markdown-vault-as-AI-memory angle is the part I actually trust here, since it stays as real files on disk instead of a proprietary store. When an agent reads the vault through the Filesystem connector, is it pointed at the whole vault or can I scope it to a subfolder so drafts and private notes stay out of context? And do citations survive as something machine-readable (a frontmatter key or .bib), or are they markdown links the model has to re-parse every time?
note.md
@hi_i_am_mimo
Great catch. Right now in the markdown files its only markdown links for internal routing (notemd://cite/*) but the name of the source alone has been enough in our tests that modern AI models like Opus 4.8 have been able to backtrack this to the right source almost every time.
There is more to come and work on regarding the LLM Memory. The core Idea where this originated from was leveraging Agents to document experiments and model training directly to the vault so the researchers can focus on experimenting and the Agent does the documentation.
Curious how the LLM memory holds up as a vault gets large. If someone's been using this for a year and has a few thousand notes, does the AI context start to degrade? Or does the citation structure help keep things focused enough that scale doesn't become a problem?
note.md
@schott_taylor
As of now note.md is not scalable enough for huge vaults to be used as LLM memory without a certain drop in quality.
We do have some interesting feature ideas on our roadmap that needs to be evaluated.
So as of right now its just a folder of .md files, but there is more to come in the future!