Reading Room Labs builds a reading room for you and your AI assistant. Shelfmark catalogs the documents and email you already have into topics you define. Flyleaf is a Markdown wiki where you and your agents write down what you learn, and keep it current.
Organized by you. Maintained by agents. Every answer cited.
Vendor renewals
The Northwind agreement renews automatically on 1 March 2027 unless notice is given 60 days before1.
Updated by renewals agent · approved by you
Northwind vendor agreement
§4 Renewal and termination
01 · The problem
The documents, threads and notes that matter live in a dozen systems. An assistant searching them on the fly sees fragments, not structure.
The memory assistants build in the background is made of embeddings and inferred links. Over time it connects things that don't belong together, and nobody can say why a document came back.
Briefs, plans and answers worth keeping end up in chat history, where the next conversation and the next agent can't build on them.
02 · The idea
Andrej Karpathy described a pattern he calls the LLM Wiki. Instead of making an AI rediscover your documents from scratch every time you ask a question, let it build a wiki: written once, kept current, and richer with every source you add. We think he's right. We also think a wiki like that needs a library behind it.
“The tedious part of maintaining a knowledge base is not the reading or the thinking — it's the bookkeeping.”
So we built a library to do the bookkeeping.
Your documents stay where they are, unchanged. Shelfmark catalogs them: a table of contents, a summary of every section, and a card on the right shelf. It keeps a snapshot, so an entry stays useful even if the original is deleted.
Out of print, never out of reach.
What you work out with your assistant gets a lasting home in plain Markdown pages, with agents that keep them current from your shelves and your tools. Every edit is versioned, attributed and comes with a reason. Scheduled agents wait for your approval before they write.
Pages have agents. Agents have skills.
The structure is yours. You define the topics; Shelfmark suggests where things belong and asks when it isn't sure. Agents and skills are short Markdown you can read and edit. Nothing about how your knowledge is organized is buried in an embedding.
Taxonomy first.
The pattern, and where it lives here
Add.
Pick a doc in Drive, a thread in Gmail, or just ask your assistant. Shelfmark files it under the right topic, and anything it isn't sure about waits in To be shelved for you.
Ask.
Ask your assistant to become an expert. It reads the way you'd use a library: catalog, then table of contents, then only the chapters it needs. Every passage it quotes links to the original and the date of the snapshot.
Keep.
Good answers become Flyleaf pages, and they go back on the shelf marked derived: context, not authority. When a source changes, your citations say so, and your agents can catch up.
Every answer comes with a call number.
Similarity search rediscovers your knowledge on every question and can't tell you why it found what it found. Reading Room Labs finds things the way you would: by where they were shelved and the words they contain. Every result has a reason you can read.
Designed to be operated by agents. Built to be understood by humans.
03 · Shelfmark · the catalog
Shelfmark is a personal knowledge base for people who work with an AI assistant. It catalogs the documents and email you already have into topics you define, and lets any assistant become an expert on one of them, citing the exact sections it read.
A short conversation, in the web app or right from Claude or ChatGPT. Your assistant proposes topics; you shape them.
Docs and PDFs from Drive, threads from Gmail. Or just ask: “add the proposal linked in this email to my shelf.”
A table of contents, section summaries and a summary card, filed under one primary topic. When it isn't sure, it asks.
“Become an expert in Project Omega.” Your assistant reads the cards, then the contents, then just the sections it needs, and cites every one.
become_expertProject Omega · 6 summary cardsget_outlineNorthwind vendor agreementread_sections§4 Renewal and terminationIt renews automatically on 1 March 2027. To stop it, notice is due 60 days before, by 31 December 2026.1
1 Northwind vendor agreement, §4 Renewal and termination
Google Doc · snapshot 24 Oct 2026 Current
04 · Flyleaf · the wiki
Flyleaf is a fast, lightweight Markdown wiki for writing down what you learn. Write pages yourself, ask the chat beside each page, message the Discord bot, or work from your own assistant. Agents attached to a page keep it up to date.
An agent is a short Markdown definition attached to a page. It can run on a schedule, when the page changes, or when an event arrives. “Every morning, update this page with my open Linear issues.”
Connect Linear, GitHub and Google Workspace through each tool's own sign-in. Skills say which tools an agent may use, and Flyleaf won't create an agent whose skills need a tool you haven't connected.
Scheduled and change-triggered agents propose their edits as a diff you accept or reject. Every save is a version, signed with who made it and why.
---
name: renewals
description: Keep the renewals table current; flag what's due in 90 days.
trigger: schedule
schedule: "0 8 * * 1-5"
skills: [linear-issues]
shelves:
- { shelf: Projects, topic: Omega › Vendors }
confirm_before_write: true
---“Content stays current because an agent keeps it that way, not because someone remembered to update it.”
05 · Better together
Each product stands alone. With one Google sign-in and one switch, they work as one reading room.
Shelfmark shelves your sources.
Flyleaf's chat and agents read your shelves and cite exact sections.
The answer becomes a page.
Add to shelf files the page as a derived source.
Citations show when a source has changed.
And round again: every page you file back makes the next answer better.
06 · Principles
You define the topics. AI files and suggests; the structure stays visible, editable and yours.
Tables of contents and keywords, not similarity scores. Every result has a reason.
Shelfmark points to your originals instead of replacing them, and keeps a snapshot in case they disappear.
AI-written material lives on a real page, and when it's shelved it's marked derived: context, not authority.
We don't build another chatbot. Claude, ChatGPT and Codex connect over MCP.
Shelves are always private. Pages are private until you share them. Nothing is used to train models.
“Markdown is the lingua franca of humans and LLMs.”
Integrations
Dashed items are on the roadmap.
07 · Trust
Each product is one service, one web app and one worker. Your content is plain Markdown and JSON in its own folder, easy to read and easy to export.
No shared content database: every shelf and every site is its own folder. Shelfmark reads only the Drive files you pick and only the email you hand it. No model is trained on your data.
Agents are limited by the platform, not by their prompt. A run can write only to its own page, holds no credentials, and stops after 15 minutes.
Coming nextEach run in its own isolated micro-VM.
08 · Who it's for
Every client's documents on its own shelf, and briefs that cite them.
Strategy, decks and investor threads your assistant actually knows.
Papers and notes cataloged by topic, and literature pages that stay current.
Specs, backlog and architecture decisions your coding agent reads before it starts, and agents that keep them in step with Linear and GitHub.
Teams next: shared shelves that respect Drive permissions, and shared wikis.
Shelfmark and Flyleaf are in early access. Join the waitlist and we'll let you know when there's a seat for you.
We'll only use your email to tell you about Reading Room Labs. No spam, ever.