Salmex I/O.

01 / The idea

An assistant with continuity.

I wanted an assistant that could remember earlier conversations and carry out useful work across tools, rather than starting over with each session.

Salmex ran on the user’s computer, with locally stored context and a choice of model providers. External model and search APIs could be enabled when needed; Ollama supported local inference.

02 / The product

Remember, act, and follow through.

The desktop app brought the assistant’s activity, memory, and scheduled work into one place. It could also communicate through Telegram, Slack, the command line, and the web.

• Context across conversations

  Persistent memory combined search and retrieval so the assistant could draw on earlier work.

• Tools and scheduled tasks

  Web search, browser automation, coding tools, and plugins let it perform tasks and run work on a schedule.

• Checks before sensitive actions

  A review system assessed tool calls and could ask for human approval before proceeding.

03 / Engineering

A system with clear responsibilities.

I designed and implemented the Go backend and Tauri/SvelteKit desktop app. The architecture separated communication channels, routing, agent execution, safety review, memory, plugins, scheduling, and the user interface.

That separation made it possible to add tools and model providers without rebuilding the rest of the application. The architecture documentation describes the individual layers and how they communicate.

04 / My role

From product idea to implementation.

I led product discovery, technical architecture, full-stack implementation, and early go-to-market work. The project was independently built, without venture funding.

My work on Salmex I/O ran from October 2025 to April 2026. The project is presented here as a concluded venture.