AI news and explainers at 7 AM weekdays, plus a Sunday weekly at 8Get it in your inbox

METAL LAB

tinyhumansai/openhuman

36,611RustGPL-3.0

An open-source desktop AI that remembers your whole digital life and runs teams of agents for you

OpenHuman is a Rust-based, open-source personal AI assistant that keeps a local, editable memory of your emails, docs, and accounts, then uses that memory to run and coordinate multiple AI agents on your behalf. It syncs your connected accounts every 20 minutes and compresses everything into readable Markdown files instead of an opaque database. It also lets an agent draft automations for you to review before they run, and reaches you through 17 messaging channels including native email.

What it does

  1. Builds a 'Memory Tree': your synced data (email, docs, chats) is compressed into scored Markdown files stored locally in SQLite and mirrored as an editable Obsidian vault, instead of a hidden vector database.
  2. Auto-fetch pulls new data from over 100 OAuth-connected accounts, 5,000+ MCP servers, and 90,000+ Skills every 20 minutes, so the agent has current context without a long learning period.
  3. TokenJuice compresses tool output before sending it to the language model, cutting token usage by up to 80% for the same information, which the project says makes a memory this large affordable to run.
  4. Acts as an orchestrator rather than a single chatbot loop: it runs checkpointed 'agent graphs' via the open-source tinyagents project, spawns sub-agent fleets up to three levels deep, and lets agents talk to each other over Signal-protocol end-to-end encryption.
  5. Workflows: the agent proposes an automation as a visual, reviewable graph (via the open-source tinyflows project) that you approve before it runs on schedules, webhooks, or channel events; a one-switch Privacy Mode keeps all inference on-device when enabled.

Why it matters

Most AI assistants start with no knowledge of your work and need weeks of chat history to become useful, while their memory and reasoning stay locked in an opaque cloud service. OpenHuman's approach — local, human-readable memory plus a reviewable multi-agent orchestrator — matters for anyone who wants an AI helper that is both immediately useful and auditable rather than a black box.

Terms in this repo

  • Memory Tree · 사용자 데이터를 요약해 점수 매긴 마크다운 파일 구조로 저장하는 로컬 메모리 시스템
  • MCP servers · AI 에이전트가 외부 도구/서비스에 연결할 때 쓰는 표준 프로토콜 서버
  • TokenJuice · 모델에 보내기 전 도구 출력 데이터를 압축해 토큰 사용량을 줄이는 기능
  • agent graph / checkpointed graph · 에이전트의 작업 흐름을 그래프로 표현해 중간에 멈췄다 재개할 수 있게 만든 실행 방식
  • Privacy Mode · 켜면 모든 AI 추론이 기기 밖으로 나가지 않도록 강제하는 설정

Repository description (English)

Your Personal AI super intelligence. A brain that builds a local-first memory of your life, a fantastic orchestrator of agent fleets and workflows, and a deep researcher.

Open on GitHubProject site

Coverage

Trending repos

All repos →

Latest from METAL LAB