Research · product report
Astro, measured against itself.
Codger got a benchmark because comparable projects existed to run it against. Astro doesn't get one here either, not because nothing compares to it, Obsidian and Logseq plainly do, but because no real head-to-head has been run yet. What exists instead is a before-and-after on a real vault: what v0.1.0 got wrong, and what v0.1.1 actually fixed, counted rather than described.
← Back to the product pageGoal
The graph isn't a map of the vault. It is the vault.
Most graph views in note apps are read-only: a picture generated from files that are really edited somewhere else. Astro's bet is that the 3D graph can be the editing surface itself, click a star to open the note, drag it to move it, shift-drag between two stars to link them, while every note underneath stays a plain Markdown file and every canvas a plain JSON file on disk. That bet only means something once real files, not the six invented demo notes it ships with, get opened inside it. So the question this page answers is narrow: on one real vault, did the architecture hold, and what broke when it didn't?
Method
Open a vault Astro wasn't built around.
The test vault wasn't written for Astro. It was an ordinary Obsidian vault: 364 notes spread across many project folders, each keeping its own Status.md, and 1,953 wiki-links written the way Obsidian writes them, as paths rather than bare titles. Nothing about it was staged to flatter or break the app. It was simply opened, and the count of what appeared and what resolved was taken before and after the fix.
Result
v0.1.0 identified notes by title. That was the bug.
A vault that repeats a filename across folders, which a Status.md-per-project habit does constantly, collapsed every note sharing that name onto one identity. Only the last one written survived on screen; the rest existed on disk and simply never appeared, with nothing in the app saying any were missing.
Notes that appeared in the app, out of the vault's 364, before and after v0.1.1 keyed notes by path instead of title.
v0.1.1's fix is structural, not a patch on the symptom: notes are now identified by their path, so two files named Status.md in different folders are two notes, not one. The same release fixed two adjacent problems the same vault exposed: wiki-links written the Obsidian way, as paths, now resolve (1,941 of 1,953 on this vault), and Constellation Mode draws a distinct figure per constellation instead of repeating a single one across the sky. A large graph also stopped rendering as a white blob, because star brightness now scales with how dense the graph is rather than only how fast the machine is.
Constraints
What ships opt-in, and what doesn't ship yet.
A vault is a folder: notes as Markdown, canvases as JSON, readable by any editor without Astro running and with no database in between. AI is entirely opt-in, a key you provide or a local runtime you point at, stored in the vault's own settings and sent only to the provider chosen. Nothing about opening a vault requires an account or a network call. What isn't built: cloud sync, so the same vault on two devices is still a manual copy, and mobile, which doesn't exist in any form.
Limits
Where this is a self-report, not a benchmark.
- There is no benchmark against Obsidian, Logseq, or Roam, the apps a vault like this is actually competing with. This page reports what broke and got fixed on one vault, not a comparative result.
- n=1 vault. 364 notes and 1,953 wiki-links is one real test case, not a sample. The exact ratios (66%, 1,941 of 1,953) are this-vault-this-run, not a general claim about every vault shaped this way.
- The identity-collision bug was specific to vaults that repeat a filename across folders, an ordinary Obsidian habit (a Status.md inside every project). A vault that never repeats a filename would never have shown the v0.1.0 failure at all, which is a limit of the test, not evidence the fix generalizes further than it does.
- 12 of 1,953 wiki-links still don't resolve in v0.1.1. That remainder hasn't been root-caused publicly yet.
- Retrieval quality, whether a note written today still answers a question in three months, is unmeasured. That's the memory-and-retrieval research track, and it needs a vault that has actually aged, not a fresh import.
- The installer is unsigned, so Windows SmartScreen warns on first run. macOS and Linux builds aren't published; they exist only as something you can build from source.
Code
Open source, and its two skills.
Astro is MIT licensed. Two skills ship inside it, astro-brain and stardust, each with its own report because neither has a comparable project to be measured against either.
Next
What would turn this into a result.
A before-and-after on one vault says the fix worked, not that Astro is a good vault to use compared to what people already run. That needs a real comparative benchmark against Obsidian or Logseq on shared tasks, and it needs the memory-and-retrieval track: the same vault, aged for months rather than imported once, queried by an agent instead of counted by hand.
The research tracks →