OpenHuman: An AI Agent That Knows You in 20 Minutes with a Karpathy-Style Knowledge Base

OpenHuman is a new AI agent that has taken the GitHub Trending list by storm, amassing over 9,000 stars in a short time, with a daily increase of a thousand stars. Unlike previous agents like "Xia" and "Ma" that require users to teach them, OpenHuman proactively learns about you. It connects to 118+ services including Gmail, GitHub, Slack, Notion, and calendar with a single authorization—no need to generate API keys manually. Once connected, its core engine polls all connected accounts every 20 minutes, fetching new emails, schedule changes, code commits, and document updates automatically.

The collected data is cleaned, compressed into Markdown fragments under 3000 tokens, and then scored, summarised by topic, timeline, and related entities to form a memory tree. This tree is stored as a local SQLite database, but also generates .md files compatible with Obsidian, allowing users to open and edit the agent's memory directly. Additionally, a feature called TokenJuice compresses HTML to Markdown, shortens long URLs, removes non-ASCII characters, and deduplicates redundancy before feeding to the LLM, cutting token consumption by up to 80%.

OpenHuman also features a Mascot—a talking virtual avatar that can join Google Meet meetings as an independent participant, taking notes and executing tasks in the background. Even without direct interaction, the agent autonomously loads to-dos and recent memories to decide what to do next. Compared to Claude Cowork, OpenClaw, and Hermes Agent, OpenHuman excels in ease of use, cost, memory, third-party integration, and automatic data synchronization. It solves three pain points: scattered API keys, fragmented data across platforms, and bloated context windows. By unifying all services into one account and running silently in the background, it offers a personalized, always-updated AI assistant that truly understands you.