Orionfold Cortex

Orionfold Cortex is a second brain for your own notes, running on your own desktop. It grew out of Orionfold Arena, where it started as the cockpit’s memory pane. Point it at a folder of notes and it builds an index: a fast lookup table that lets you ask questions in plain words and get back the right passages, with a citation for each one. All of it runs on the machine under your desk, and none of it leaves.

Why it exists

An index of your notes is the easiest thing to let rot. You build it once, it works, and then the notes grow, a rebuild quietly drops a tenth of them, and nothing tells you. The number that matters, “does a question still find its answer,” stays invisible until the day it fails you. Cortex makes that number something you can see, chart, and gate on one screen.

What you can do

• Read the health of your memory. One screen shows how much of your corpus is indexed, how much has gone stale, and how many facts carry a stamp saying where they came from. No guessing.
• Rebuild without a script. One click queues a re-index and a scoring run behind it. You watch both move across a jobs board while the machine does the work.
• Trust the gate. Every rebuild scores itself against a fixed set of real questions with known answers. If the new index finds fewer of them than the old one did, it is flagged instead of shipped. A worse memory never replaces a better one quietly.
• Ask with a trust filter. Type a question in plain words and get cited passages back. Each one is tagged by how much you can trust its source, so a number measured on your own machine is never silently mixed up with a claim someone made on the internet.

Jargon, in plain words: an index is a fast lookup table over your notes. Recall is the share of questions where the index finds the right passage. Provenance just means “where a fact came from,” and Cortex stamps it on every piece it stores.

Honest numbers

On its first measured run, over 49 articles split into 313 pieces, Cortex found the right passage in its top five answers 40.91% of the time, and the right note 72.73% of the time, across 44 real questions. Those are the true numbers for the simple baseline it runs today, with no extra ranking step yet. They are printed on the product, not rounded up.

The first real run also caught a real bug: a missing one-line import that eight unit tests had slept through, because they all used a stand-in instead of the real code. Driving the tool through its own screen found it in minutes. That is the whole idea: a memory you operate, not a black box you hope still works.

Private by design

Your documents, the index, and your questions all stay on your machine. Only aggregate scores are ever shared, and only if you publish them.

Built on fieldkit

Cortex is a thin surface over the fieldkit toolbox: the memory module owns the index and the source stamps, the Arena control plane runs the jobs, and the same test machinery that grades our models grades the index. The parts already existed; Cortex is what you get when you point them at each other and add the one missing piece, a memory that grades itself.