Architecture
OpenHuman is built on the OpenClaw architecture and open-sourced under the GNU GPL3 license. This page explains how the major components connect.
The three pillars
OpenHuman's architecture rests on three pillars that work together:
Neocortex is the memory engine. It ingests data from connected sources, builds knowledge graphs, manages tiered memory, and provides the recall capabilities that power both conscious queries and subconscious processing. Detailed in Neocortex.
Multi-agent orchestration distributes work across specialized agents rather than relying on a single monolithic model. An orchestrator agent manages routing, personality, and context distribution. Specialist agents handle specific domains: communication analysis, document synthesis, task management, trading. Agents execute in parallel, not sequentially, enabling real-time responsiveness.
Privacy-preserving inference ensures that raw data never leaves the user's device. Data is encrypted on-device with AES-256-GCM. Encryption keys never leave the device. Only compressed metadata and summaries are processed server-side. Detailed in Privacy & Security.
How data flows
Ingestion. Data arrives from connected sources: Telegram, Slack, Gmail, Notion, blockchain wallets, and others. Each source has its own connector that handles authentication and data retrieval.
Compression. Neocortex processes raw data on-device. Semantic deduplication removes noise. Entity resolution links references across sources. Temporal weighting prioritizes recency. The output is a compressed knowledge graph, not raw text.
Storage. The knowledge graph is stored in Neocortex's tiered memory system. Raw data is discarded after compression. Only structured metadata and summaries persist.
Conscious processing. When you make a request, the orchestrator routes it to the appropriate specialist agent(s). Those agents query Neocortex for relevant context, process your request, and return a result.
Subconscious processing. Independent of your requests, the subconscious system triggers periodic memory recalls from Neocortex. These feed into a self-learning loop that surfaces proactive insights, patterns, and recommendations.
Output. Results are presented to you directly or exported to connected tools like Notion and Google Sheets. Only structured, compressed intelligence leaves the device. Raw data never does.
Model-agnostic design
OpenHuman is not locked to any single AI model. The compression engine and memory layer sit on top of the AI infrastructure, not inside it. Today the system works with specific models. Tomorrow it could feed context to any model: GPT, Claude, Gemini, Llama, Mistral, or whatever comes next.
This is a deliberate architectural choice. AI models are commoditizing. Performance is converging. The real differentiator is the context you feed the model, and OpenHuman owns the context layer.
Open source
OpenHuman is publicly available on GitHub under the GNU GPL3 license.
GitHub: github.com/tinyhumansai/openhuman Neocortex benchmarks: github.com/tinyhumansai/neocortex/tree/main/benchmarks
Contributions, feedback, and issues are welcomed. The project is in early alpha.
Last updated