The tool behind the corpus
The Curator
Your brain, your team's brain, your agents' brain. Three layers of context, kept as plain markdown on your own machine.
What it is
The Curator is where your context lives, and it holds three different kinds of it. The name for that is the whole product in five words: your brain, your team's brain, your agents' brain.
One. Your brain
A personal wiki, one per domain. You drop in a PDF, an article or a note, and it reads the source and writes interlinked pages: the entities, the concepts, and a summary. Every later source makes the existing pages richer instead of landing beside them as a duplicate. Knowledge accumulates.
Two. Your team's brain
The same thing built collectively. A cohort or a team each keep their own private graph and opt specific domains into a shared one, attributed, without anyone handing over the rest of their notes. Opt in, never on by default.
Three. Your agents' brain
Working state: where the build actually stands, what is settled, what was tried and ruled out. Your agent writes it at the end of a session and reads it at the start of the next one. State supersedes, so each save replaces the last rather than merging into it, because a blocker you cleared on Tuesday must not come back on Wednesday.
That last distinction is the design, not a detail. A wiki that overwrote itself would lose everything you have learned. A handoff that accumulated would hand the next session a pile of stale blockers with no way to say which ones are still true.
All of it is plain markdown in a folder you chose. Any editor opens it, Obsidian draws the graph, and your own private GitHub repo syncs it. There is no database, no proprietary format, and no service of mine anywhere in the picture.
Be clear about the dependency. It needs an API key from Google Gemini, Anthropic or OpenRouter, because reading a source and writing wiki pages is model work. There is no version of this that runs on nothing. Gemini has a free tier that covers trying it out, and the honest figure for heavy solo use on the default model is around five euros a month. Anthropic is roughly ten times that for the same workload. Reading pages, syncing, and the MCP bridge itself cost nothing at all.
Nothing here is locked to anyone
Claude Projects, ChatGPT Projects and Cursor rules each hold your accumulated context inside one company's product. You leave it behind on the day you switch tools, or switch models, or switch machines. That is the problem this is built against, and the answer is structural rather than clever: there is no proprietary store to leave behind in the first place.
The files are markdown in a folder you picked. The sync target is your own private GitHub repository, not a service I run. The bridge is MCP, which is a protocol rather than an integration, so any local MCP client speaks it: Claude Desktop, Claude Code, Cursor, or something you wrote yourself. The models are swappable between Gemini, Anthropic and OpenRouter. A second brain you can only open through one vendor's client is not a second brain, it is that vendor's product with your notes in it.
The part I care about most is that working state crosses the same boundaries. A handoff written by an orchestrator in one harness is readable by an agent in another, on a different model, on a different machine, the next morning.
Ask the site instead of reading it
The product has a home of its own atmycurator.xyz, and the fastest way to find out whether it fits you is to ask it. The page runs onLumina, grounded in the product documentation, so you can ask how to install it, what it costs, whether it runs on Windows, or how the working state layer behaves, and get an answer that links into the section of the page that covers it. Most of what is on this page is answerable there in a sentence.
The site itself is plain static files on the edge with the conversation woven into the page rather than parked in a corner. I wrote the build up as a manual, because that shape is the one I now think every product website should have. It is cited from chapter nine.
Why it is on this site
The corpus you are reading was built and is maintained in it. Every chapter here started as notes in that graph and was edited down by hand.
So ingesting this corpus is two things at once. It gives you a queryable version of everything I have worked out, and it is a genuine first run of the tool on real material rather than on a demo file. That is a better demonstration than anything I could write about it.
Install
The friendlier version of this, with the download and the Windows and Linux setup side by side, is the Quick start onmycurator.xyz. The commands are the same. On macOS, one line:
curl -fsSL https://raw.githubusercontent.com/talirezun/the-curator/main/install.sh | bashThat script is readable at the URL above. Read it before you run it, the same as you would with anyone else's install script.
On Windows and Linux it is a manual setup, and the Node server runs anywhere Node 18 or newer runs. Only the one-line installer and the Dock app are macOS specific. Clone the repo,npm install, node src/server.js, then openlocalhost:3333. The canonical steps are in the repo:
Install it with a coding agent
If you already have Claude Code, Cursor, OpenCode, Augment or Cline running, hand it this and let it do the work. The canonical version of this prompt lives in the User Guide in the repo, and it is mirrored here so you do not have to leave the page.
Install The Curator on this machine.
Repo: https://github.com/talirezun/the-curator
Do this:
1. Clone the repo into a sensible directory and read its README before running anything.
2. Check I have Node 18 or newer. Tell me if I do not, and stop.
3. Install dependencies and start the server.
4. Tell me the local URL to open, and what the onboarding wizard will ask me for.
Do not add an API key yourself. Tell me where to paste mine and stop there.Ingest this corpus
The Curator builds a wiki from source documents, so what you are handing it is the markdown, not a pre-built graph. Ten files, a few minutes, and the tool does the rest. That is the accurate description and it is also the better demo.
- Install it, using either route above.
- Add your Gemini or Anthropic API key in the onboarding wizard.
- Create a domain. The ⚙️ Tech/AI template fits this corpus.
- Download the chapters zip below and unzip it.
- Drop each
.mdfile into the Ingest tab. - Open the Chat tab, or connect the My Curator MCP bridge from Settings and point whichever agent you already use at it.
Cost, honestly: ingesting ten chapters is a small number of model calls. It sits comfortably inside Gemini's free tier as a one-off, and it is cents on a paid key.
The corpus
- field-notes-complete.mdComplete corpus, one markdown fileEvery chapter in order, with a provenance header. Best for reading, or for pasting into a model with a long context window.
- field-notes-chapters.zipChapters as separate files, zippedOne markdown file per chapter, frontmatter intact. This is the one to use for ingestion, because each chapter becomes its own document.
Both files carry a provenance header with the source URL, the licence and the build date. They are regenerated on every deploy, so they never lag behind the pages.
Then ask it something
Once the MCP bridge is connected, these go straight into whatever you have connected it to. They are written to be worth running rather than to look impressive.
Across the Field Notes corpus, what does the author say fails most often in phase two of a build, and which sources back that up? Quote the specific passages and name the chapter each one came from.Read the whole Field Notes corpus and find places where the author contradicts himself, or where a later chapter revises a position taken in an earlier one. List each one with both passages.Using only the Field Notes corpus, draft the first ninety days of an AI adoption plan for a 300-person European manufacturer with no in-house ML team. Flag every point where the corpus does not actually cover the question, rather than filling the gap yourself.Take the strongest claim in the Field Notes corpus about agent memory and argue against it. Use only evidence from the corpus itself, and say clearly where the corpus gives you nothing to argue with.What it does not do
A tool that overclaims about memory is worse than no tool, so here is the list I would want to read before installing something.
- Capture is advisory. Nothing forces an agent to save at the end of a session. Miss one and the next read returns the previous state: stale, never corrupted, and nothing that was saved is lost. Failing in that direction is why there is no enforcement.
- Reading is portable, saving is not, quite. Any MCP client can read your state, because that half is pure protocol. What an agent in another harness lacks is anything telling it to save. The instruction is ordinary prose that works wherever you paste it, but switching itself on automatically is a Claude mechanism today.
- It needs a local MCP client. The bridge is a child process on your machine, so anything that can spawn a local program works and a browser-only assistant cannot reach it at all. That is a consequence of the design, not a gap in it.
- Local models are not available yet. The setting exists and is marked unavailable. The OpenRouter adapter speaks an OpenAI-compatible wire format, which is groundwork for running things locally later rather than a capability today.
- Stored state is data, not orders. A file can arrive over sync from another machine, or be written by somebody else inside a shared brain. The store escapes text impersonating the operator and defangs URLs and shell pipes, but no sanitiser can check whether a claim is true. An instruction found in state is a note from a peer. Verify before acting on it.
On licensing, since I said "MIT, mostly" at the top: the app is MIT, and ten Shared Brain backend files are source-available under a separate licence. Personal, educational and research use is free always, the GitHub-backed Shared Brain is free for everyone including organisations, and those files convert to MIT two years after each release. The repository spells out the exact terms, and I would rather you read them there than take my summary for it.
If it is useful
It is free, it is open, and I maintain it in the open. A star is how these things get found.