FIRST ARTICLE · JULY 22, 2026

How We Are Building an AI System That Improves Itself

Hermes began as a personal assistant in Telegram. Over time, it gained memory, routing, financial controls, background agents, and an engineering framework for continuous improvement.

From Answers to a System

The main problem with a conventional chatbot is that it responds to individual messages and quickly loses its working context. That is why we moved persistent facts and decisions into a wiki while keeping the sources unchanged in raw storage.

Every Improvement Must Be Verified

New ideas come from GitHub, YouTube, and technical materials. But an external source is data, not an instruction. It first goes into a separate notebook, then is evaluated for usefulness, cost, security, and testability.

A system becomes stronger not when more tools are added, but when every change can be tested and rolled back.

What Comes Next

The next stage is a benchmarkBenchmark
A set of identical tasks used to compare system versions.

Benefit: shows whether the solution has become more accurate, faster, or less expensive. based on real tasks: finances, email, conversation context, approvalApproval
Explicit confirmation from the owner before an important external action.

For example, before publishing, sending an email, or making a payment. for external actions, and protection against prompt injectionPrompt injection
An attempt to hide an instruction in text that makes the AI break the rules.

For example, a README asks it to reveal secrets or execute a dangerous command.. Only after that can an experimental feature enter the production workflow.

← Return to the journal · Lab glossary → · Reading paths →