Louis Beaumont

Table of content

Louis Beaumont is building the infrastructure for AI memory — systems that capture everything you see, hear, and do on your computer, stored locally, and made accessible to AI agents.

Screenpipe: Open-Source Rewind

His flagship project screenpipe (16.5k+ stars) is an open-source alternative to Rewind.ai:

24/7 recording of screen and audio
100% local — your data never leaves your machine
OCR + transcription — searchable text from screenshots and audio
AI integration — ask natural language questions about your digital history
MCP server — connect your screen history to Claude and other AI tools

The philosophy is radical data ownership: “rewind but open source. 100% local. you own your data.”

Technical Approach

Screenpipe optimizes for minimal footprint:

~10% CPU usage
4GB RAM
~15GB storage/month (heavily compressed)
Works offline

Built in Rust for performance, with TypeScript for the UI layer. The project includes automatic PII redaction before sending context to AI.

The Mediar Stack

Beyond screenpipe, Louis ships prolifically through Mediar AI:

Computer Use Automation:

terminator — “Playwright for Windows computer use” (1.3k stars)
useterminator.com — desktop automation
autoehr.com — healthcare automation
autosap.ai — SAP automation

MCP Servers:

Background

Louis’s path to personal AI is unconventional:

French intelligence — satellite intelligence/engineering
OVHCloud — end-to-end hardware observability for data center devices
Game hacking — reverse-engineered Diablo 2 and Dofus protocols, built bot orchestration systems
Early RL work — trained human-level AI for X-Moto using reinforcement learning (2019), self-driving cars with deep RL (2018)

This progression — from reverse engineering game protocols to recording and understanding human computer use — shows a consistent interest in capturing and automating digital behavior.

Philosophy

Louis embodies the “ship fast, iterate faster” ethos. His GitHub shows 162 repos and multiple past projects that became stepping stones to screenpipe:

rewind.sh — “memory for devs”
screen-to-crm.com — sales automation
HAL-9100 — agents infra for edge defense systems

The common thread: making compute understand and automate human workflows.

Key Ideas

Memory as infrastructure — screen recording isn’t surveillance, it’s giving AI the context it needs to actually help you
Local-first is non-negotiable — for something as intimate as screen recording, data ownership must be absolute
MCP as integration layer — multiple MCP servers that connect different data sources (Obsidian, Toggl, Oura, screen history) to AI
Automation at the OS level — terminator brings Playwright-style automation to desktop apps, enabling true computer use agents