Ghost Protocol

What we built, and what we measured afterwards

Each figure on this page is either reported by the study it comes from or taken from our own engineering records. Where a number is an estimate, the page says so.

Memory is not the moat, judgment is

Client
Multi-agent operations team (anonymized)
Sector
AI agent operations
Stack
Claude · MCP · Wyrm · SQLite
Dates
2026-07-02updated 2026-07-04

The problem

Two AI agents on the same desk, comparable models. One had the richer memory and a standing instruction to act first and stop asking. In a single week it rewrote a teammate's working automation unprompted and answered the same question with confidently wrong numbers three times. Its memory logged every failure faithfully, afterwards, and changed nothing.

What we built

The other agent ran on Wyrm, which treats memory as a decision-time layer rather than a diary: it checks its record of past failures before acting, loads the relevant rule instead of improvising, reads numbers from the system of record, and pauses for a human before anything destructive. The write-up is grounded in a 108-agent research pass across 25 sources, with 21 claims verified.

Measured

Negative-learning firewall, recall / precision
100/100
Tests in the Wyrm suite
3,395
Sources reviewed for the study
25

Firewall and test figures are Wyrm's own engineering records at 9.2.3; the research figures are the study's. Company, product and personal details are withheld in the study.

Read the case study 

One operator, thirteen projects, zero dropped context

Client
Forward-deployed operator (anonymized)
Sector
AI memory, solo operations
Stack
Claude Code · MCP · SQLite · zero-knowledge sync
Dates
2026-06-24

The problem

One person running more than a dozen client and product workstreams at once, with an AI assistant as the force multiplier. The bottleneck was never capability. Every session started cold: context re-explained from scratch, decisions evaporated between sessions, the same dead ends retried days apart.

What we built

Wyrm as the persistent memory under Claude Code: an MCP server backed by SQLite on the operator's own machine, loaded automatically at the start of every session, so the assistant picks up the open threads and the recorded failures without a re-brief. One memory across Claude Code, Copilot, Cursor, Windsurf and Codex.

Measured

Sessions held
77
Threads resumed
257
Projects in one memory
13

Read directly from the operator's own Wyrm memory by their CLI assistant, the way Wyrm is used every working day. Identifying details withheld.

Read the case study 

How we digitized a traditional restaurant in 2 weeks

Client
Upali's, Colombo
Sector
Restaurant, food and beverage
Stack
2025 build: PHP 8 · MySQL · PWA · PayHere. Rebuilt on Cloudflare Workers · D1 · KV · R2
Dates
Study 2025-01-25Cloudflare rebuild: later, undated here

The problem

Handwritten orders, phone-based delivery, printed menus in one language, and a 25 to 30 percent commission on every Uber Eats or PickMe order. For a restaurant on thin margins that is the difference between profit and loss.

What we built

A restaurant platform built from scratch in two weeks: online ordering with delivery, pickup and dine-in, a kitchen display, a 21-module admin, an installable PWA, PayHere payments plus cash on delivery, and 103+ menu items in English, Sinhala and Tamil. The platform has since been rebuilt on Cloudflare Workers with D1, KV and R2, keeping the three languages and adding time-of-day menu pricing, with hosting at zero dollars a month. The study's figure of roughly $18,000 a year kept from commission is an estimate, not a measurement.

Measured

Monthly hosting after the Cloudflare rebuild
$0
Commission on the restaurant's own orders
0%
Languages: English, Sinhala, Tamil
3

Hosting and language figures are from the Cloudflare rebuild; the commission figure is from the 2025 build the study describes.

Read the case study 

Measured, not yet written up

Work with a number against it and no study yet. Each figure is from our own engineering records for that deployment; the write-ups follow when the clients have signed off on what can be shown.

  • 2

    DragonLair

    Hotel operating system · hotel tenants

    Multi-tenant hotel OS live on Cloudflare Workers. A second hotel provisioned on the same deployment; guest portal, billing, and a double-booking guard.

  • 23

    Winfield Sports

    Operations platform · tests

    A checklist outage found and fixed; materials moved into D1; site photographs on R2 behind authentication; 23 tests around the fixes.

  • 7.0 ms

    Ceylon Seaweeds

    Site · render time, from 13.9 ms

    Render time cut from 13.9 ms to 7.0 ms, holding 57.5 frames per second.

  • < 2%

    ITP

    WordPress LMS to Cloudflare · visual difference

    A WordPress learning platform mirrored onto Cloudflare: 75 pages at under 2 percent visual difference from the original.

  • PhantomDragon AI

    Security engagement · in progress

    An AI-driven security engagement compressed from weeks to days. The full study is in progress; no figure is published until it is.

Talk to an engineer

Email ryan@ghosts.lk or message +94 71 055 5055 on WhatsApp. Ryan replies within 24 hours. Working hours are Monday to Friday, 09:00 to 18:00 Colombo time (UTC+5:30).

Book a 15-minute call Buying for a company? Procurement details