OpenHuman's Markdown-tree memory is a genuinely different architecture for local agents

OpenHuman is a local-first agent harness built on three interlocking ideas: persistent memory, durable multi-agent orchestration, and deep research tooling — all installed without touching a config file.

The memory architecture is the most distinctive piece. Rather than a vector store, OpenHuman compresses your connected accounts — email, calendar, repos, messages — into scored Markdown trees stored in SQLite, then mirrors them as an Obsidian vault you can open and edit directly. An auto-fetch loop pulls fresh data every 20 minutes, so the agent's context is continuously updated rather than frozen at onboarding. The README credits Andrej Karpathy's LLM knowledgebase concept as the direct inspiration. A token-compression layer called TokenJuice is what makes this practical: the README claims up to 80% fewer tokens from tool output, which is the only reason a memory this broad doesn't become prohibitively expensive at inference time.

The orchestration layer runs on two open-source sub-projects the team also maintains: tinyagents handles checkpointed graph execution (agents pause for human approval, survive restarts, resume mid-run, and return root-cause reports when stuck), while tinyflows powers the visual workflow canvas. The agent proposes an automation, you review it on a canvas, and the saved workflow fires on schedules, webhooks, or channel events. That's a meaningfully different model from most harnesses, which either run a single loop or require you to write the automation yourself.

Agent-to-agent communication runs over Signal-protocol end-to-end encryption with x402 USDC payment rails, so multiple OpenHuman instances — or third-party agents like Claude Code or Cursor — can be orchestrated together without any server seeing plaintext. The README is explicit that keys never touch disk.

The integration surface is wide: over 100 OAuth connectors, more than 5,000 MCP servers, and a claimed 90,000-plus skills. Seventeen messaging channels are supported natively, including IMAP IDLE and SMTP for email. There's also a meeting-join capability covering Google Meet, Zoom, Teams, and Webex with live transcription — something none of the three competitors named in the README's comparison table offer.

Privacy Mode is enforced in the Rust core: one switch, and no inference leaves the machine. The build stack is Node.js 24+, pnpm, Rust 1.93.0, and Tauri with a CEF shell, which signals a desktop-native architecture rather than an Electron wrapper.

The project is in early beta by its own admission, and the README is candid about rough edges. The comparison table is self-authored and should be verified against each competitor independently. The AGI framing in the footer is marketing noise. What's actually here is a well-specified local agent harness with a memory model that's genuinely different from chat-scoped or plugin-reliant alternatives — and that difference is architectural, not cosmetic.