Ari’s Space*   est. 2026
Skip to content
ari*s space[ a. ]
← Libraryworkflows / 2026-08-06

I packed the library into an ICM you can hand an agent

A 6 August 2026 snapshot of 64 public library units, organised as an ICM knowledge bundle: small routing files, topic shelves and cards that point to the live essays.

7 parts / Hover to explore. Click to read.

A 6 August 2026 snapshot of 64 public library units, organised as an ICM knowledge bundle: small routing files, topic shelves and cards that point to the live essays.

CLAUDE.md -> CONTEXT.md
                 |
          topic index
                 |
          disposition
                 |
        relevant episode cards
                 |
       ranked reading links
                 -> live essays
The dated bundle routes a question to a few cards and then to the full public essays. It is not an automatically updating mirror or an offline copy of the whole site.

The library works well when you want to browse. An agent has a different job: understand your question, find the most useful essays and give you links with a reason to open them.

So I made another form of the same public knowledge. A plain folder with a catalogue, small routing files and one card per unit. The full essays stay on the site. The folder helps you find them.

↓Download the knowledge bank (zip, free, no gate)

Give the agent a library it can walk

ICM, Interpretable Context Methodology, is the folder-based method published by Jake Van Clief and collaborators. Structure carries part of the agent's working process: numbered shelves, small routing files and Markdown as saved state.

For this corpus I used its knowledge bundle form. The goal is to make the teaching navigable, rather than paste every essay into a conversation.

aris-space-icm-knowledge-bank/
├─ CLAUDE.md              who you are + how to answer
├─ CONTEXT.md             one-screen map of the shelves
├─ AGENTS.md              pointer for non-Claude seats
├─ _index/                catalog (log + by-topic)
├─ _meta/schema.md        card contract
├─ _shared/advice-voice.md
├─ corpus/_index.md       what was harvested
└─ bundle/
   ├─ voice/              always-load orientation
   ├─ dispositions/       routers by kind of ask
   └─ episodes/           one card per deep dive / lesson

The routing files point to the material. They stay small because they do not also contain the books.

Each episode card has a title, a live URL on its own line, one paragraph of context, guidance on when to recommend it and links to related cards. The agent opens a few relevant cards, ranks them and points the reader to the site. It should not invent essays to complete a tidy list.

Catalog points. Cards excerpt. The URL is the book.

What the download contains

The supplied zip is the 6 August 2026 snapshot: 64 units, covering long-reads, how-tos, courses and lessons. It is not a claim that today's whole library contains 64 items.

Topic shelves include workflows, tools, dispatch, memory, workspaces, ARI-OS and doctrines. Disposition files route common questions about folders and loops, orchestration, memory, tools, outcomes, workspaces, dispatch and courses.

The answer format is part of the method:

## Best match
**Title**
https://aris-space.com/documents/...

One paragraph: what it teaches and why it fits.

## Top three
1. **Title** - URL
   One short paragraph
2. ...
3. ...

Give one best match and three ranked choices, with a short explanation for each. The best match can also be number one in the list; say so. The point is to help someone choose what to read, not to make them sort twenty links themselves.

Three ways to use it

1. Open it with a local agent

Unzip the folder and open it in Claude Code, Codex, Cursor or another agent that can read local files. Ask a question with an actual problem in it:

I keep losing work in chat history. What should I read? Best match and top three with links.

The intended reading order is CLAUDE.md, then CONTEXT.md, then _index/by-topic.md, a relevant disposition and one to five episode cards. AGENTS.md points non-Claude seats towards the same instructions. The output should be ranked recommendations with URLs.

2. Browse it yourself

Open _index/by-topic.md, choose a shelf and read three cards. Follow the URLs when one fits. The folder should be understandable without an agent; that is a useful test of whether the structure is doing its job.

3. Use a cloud ICM desk

The August launch coincided with Jake Van Clief, owner of the Clief Notes community, opening a cloud workspace platform. The account at the time described uploading ICM folders and second brains, then working on them together.

The described setup used one strong model with a working harness, a separate container for each workspace, rendered Markdown, file editing, package installation, Python and Playwright, and shared spaces for multiple people. Access was rolling out to VIP and alpha users while he covered the AI cost of testing.

Those are the August rollout details, not a promise of current access, pricing or unchanged features. The original routes were:

The bank was made to fit that moment: a finished public teaching folder you could put on a desk built around files. The same zip, CLAUDE.md and best-match/top-three contract work as the starting material locally or in a cloud workspace that supports them. Moving a folder does not grant a host tools or permissions it lacks.

Keep the bank's job small

This is a librarian for the public corpus. It offers context and directions to the full argument.

It does not replace the live graph at /documents. It is not a vector database, a multi-agent framework, private operational material or an offline copy of the whole site.

The card helps you decide whether to open an essay. The essay does the teaching. Keeping those jobs separate makes the answer easier to use.

  ASK                         BANK                         SITE
  ---                         ----                         ----
  "help me with X"   -->   best + top 3 + URLs   -->   full deep dive
                           one paragraph each           when they click

Keep the snapshot in date

The cards were harvested from the site's published content tree. They are a product of that harvest, not an automatically updating mirror.

When the library changes, regenerate the bundle and replace the zip after checking its contents and links. Until then, the download remains the dated snapshot described above. The live essays are the source of truth, including where their wording has been revised since a card was made.

Try it

Download the public zip, unzip it and ask a file-reading agent for a best match and top three. Or open the topic index yourself. If you use a cloud desk, upload only the public folder when you have access and want to use that service.

Start with a question you already need help answering. The useful result is a short reading route you can follow.

//A<3

[ comments ]

guests welcome. members show first in the list

no comments yet. start the thread_

1 / 9