GitHub - tinyhumansai/openhuman
OpenHuman is a local-first AI assistant that combines persistent memory, multi-agent orchestration, and deep research capabilities into a single personal AI system It uses a "Memory Tree" architecture that compresses user data (documents, emails, chats) into scored Markdown trees stored in SQLite, mirrored as an editable Obsidian vault—avoiding vector-database black boxes The system features a split-brain architecture with a fast reflex agent for triage and a deep reasoning core that delegates t
Analysis
TL;DR
- OpenHuman is a local-first AI assistant that combines persistent memory, multi-agent orchestration, and deep research capabilities into a single personal AI system
- It uses a "Memory Tree" architecture that compresses user data (documents, emails, chats) into scored Markdown trees stored in SQLite, mirrored as an editable Obsidian vault—avoiding vector-database black boxes
- The system features a split-brain architecture with a fast reflex agent for triage and a deep reasoning core that delegates to worker fleets, powered by the open-source tinyagents framework
- TokenJuice technology compresses tool output by up to 80% before it reaches the LLM, dramatically reducing inference costs for memory-heavy workloads
- OpenHuman includes 100+ OAuth integrations, 5,000+ MCP servers, 90,000+ Skills, model routing across cloud/local providers, and an agent economy with Signal-encrypted agent-to-agent orchestration and USDC bounties
Why It Matters
OpenHuman represents a significant architectural shift toward personal, local-first AI agents that can rapidly onboard user context without the typical cold-start problem plaguing most AI assistants. Its Memory Tree approach offers a transparent, editable alternative to opaque vector-soup systems, while its split-brain design and token compression make sustained multi-agent orchestration economically viable for individual users.
Technical Details
- Memory Architecture: Uses Karpathy-style LLM Knowledgebase principles—auto-fetch pulls data every 20 minutes from connected accounts, then Memory Trees compress and score the content into Markdown files stored in SQLite, mirrored as an Obsidian vault for human-editable access
- Split-Brain Agent Design: A fast reflex agent handles inbound traffic triage while a deep reasoning core delegates to worker fleets, all orchestrated on durable checkpointed graphs via the open-source tinyagents framework with per-call cost tracking
- TokenJuice Compression: Tool outputs are compressed before entering the model context window, achieving up to 80% token reduction while preserving information fidelity—critical for affordability given the system's massive memory footprint
- Integration Ecosystem: 100+ OAuth integrations (Gmail, Notion, GitHub, Slack), 5,000+ MCP servers, 90,000+ Skills, plus managed Exa web search, native Whisper voice, browser automation, and model routing across cloud APIs and local Ollama models
- Agent Economy: Implements @handle identities on tiny.place with Signal-encrypted agent-to-agent orchestration and x402 USDC bounties for microtask trading, with cryptographic keys never touching disk
Industry Insight
- The rapid onboarding claim ("gets to know you in minutes") directly addresses the #1 friction point in personal AI adoption—most agents require weeks of context building before becoming genuinely useful, and OpenHuman's auto-fetch + Memory Tree pipeline could become a standard pattern
- TokenJuice's 80% compression ratio demonstrates that pre-model output compression is a viable cost-reduction strategy for memory-intensive agent systems, likely to inspire similar approaches across the ecosystem
- The combination of local-first privacy (Privacy Mode enforced in Rust) with cloud model routing flexibility positions OpenHuman to capture the security-conscious segment that demands both data sovereignty and access to frontier models
Disclaimer: The above content is generated by AI and is for reference only.