# Salmex I/O

> Archived. Salmex I/O was a personal AI operations platform that ran locally — persistent memory, multi-channel reach, agentic execution, and human-in-the-loop safety. It has been sunset and is no longer available for download or sign-up. This site remains as a record of what was built.

Salmex I/O was a self-hosted AI operations platform built as a single Go binary. It connected to any LLM provider (Anthropic, OpenAI, Gemini, Ollama), stored everything in PostgreSQL with pgvector, and acted on your behalf across Telegram, Slack, Discord, CLI, and web — with a 4-tier LLM judge that escalated high-risk actions to you before execution.

Built independently in London. The product has since been sunset and is no longer maintained.

## Status

- Status: Sunset / archived. No longer distributed, sold, or supported.
- Downloads and sign-up: removed. No new licences are being granted.
- This website: kept online as a record of the work and its architecture.

## Key Facts

- Category: Personal AI operations platform / AI agent framework
- Language: Go (single compiled binary, no runtime dependencies)
- Database: PostgreSQL + pgvector
- LLM Providers: Anthropic (Claude), OpenAI (GPT-4, o3), Google Gemini, Ollama (local)
- Search Engines: Perplexity, Brave, Google (intent-based routing, 2-tier cache)
- Channels: Telegram, Slack, Discord, CLI, HTTP, web
- Architecture: 8-layer (Agent → Channel → Gateway → Plugin → Memory → Judge → Scheduler → UI)
- Security: AES-256-GCM encrypted config, LLM judge with 4 risk tiers, subprocess-isolated plugins
- Memory: Hybrid vector + BM25 retrieval, extraction pipeline, confidence decay, 3-tier (session/working/long-term)
- Creator: Salmen Hichri (https://salmenhq.com/)

## Features (as built)

- Persistent long-term memory with hybrid vector + BM25 retrieval and reciprocal rank fusion
- Multi-channel communication — same agent, same memory across Telegram, Slack, Discord, CLI, web
- Multi-provider LLM support — switch between Anthropic, OpenAI, Gemini, Ollama without losing context
- Human-in-the-loop safety — LLM judge reviews every tool call, 4 risk levels, real-time approval via preferred channel
- Embedded coding agent — read, write, edit, execute, grep, find files; session trees with branching
- Multi-engine web search — Perplexity, Brave, Google with intent-based routing and deep research mode
- Plugin system — JSON-RPC 2.0 over stdio, crash recovery, lazy start, health checks
- AES-256-GCM encrypted configuration — hot-reload, change history, rollback, no plain-text secrets
- Scheduled tasks — cron expressions, natural language scheduling, dead letter queue
- Session branching and compaction — tree-structured conversations, iterative summarisation
- 40+ typed events — structured event bus, SSE streaming, JSONL transcripts
- Full local operation — your data never left your machine unless you chose to connect external APIs

## Links

- Home: https://salmex.io/
- Sunset notice: https://salmex.io/sunset
- Blog: https://salmex.io/blog
- The Case for AI You Own (essay): https://salmex.io/blog/the-case-for-ai-you-own
- Architecture: https://salmex.io/docs/architecture
- Licence (archived): https://salmex.io/licence
- Extended info: https://salmex.io/llms-full.txt