Horizon Context, the memory a company and its agents think with

The problem was not information, it was trust

Everyone had context. Business in emails, ads in dashboards, design in Figma comments, engineering in the code and the tickets. None of it talked to the others, and the cost of that grew with every new site. The first version, built in the spring, stalled within two months: an inbox nobody dispatched, status files overwritten by automation.

The second version started from a different question. Not how do we store what we know, but what would it take for a person and an AI to act on it without checking with me first.

One memory, two layers

The method is written once for every site: how the group builds, sells, pays, supports, complies. Then the activities, the companies, and one entry per product that only says what differs from the method. Decisions are logged append-only. Anything unfiled lands in an inbox with a seven-day rule.

Every file carries a layer. Business keys, company names and legal data live in the internal layer. The product layer is what a customer of that site may know. A check runs on every change and blocks any company name or forbidden term from the product layer, so the support assistant that reads it cannot leak what it never sees.

horizon-context/
├── horizon-web/      # the method, written once
├── activities/       # what each kind of service is
├── companies/        # one file per legal entity · always internal
├── sites/            # one folder per product, two layers
├── people/           # who works on what, dailies
├── initiatives/      # internal projects
├── tools/            # the check on every change
└── _inbox/           # unfiled, dispatched in seven days

The shape of the memory. Internal names stay inside it; this is the structure, not the content.

The rules agents work under

A fact without a source does not go in. When sources disagree, the agent does not arbitrate: it writes both versions, dated, and a person decides. Agents extract facts and propose changes; I review and merge. The team never has to write, only correct. Nothing is deleted, nothing is overwritten, and no secret ever enters the memory, only where it lives.

How I know it works

A blind quiz: twenty questions given to an agent that had never seen the work, answering from the memory only and citing its source. Sixteen right, one corrected, zero false, and three the team itself could not answer. Then the CEO and the CTO reviewed the summary statement by statement, and their corrections went back in the same day. That loop, test blind, correct, test again, is the product.