OpenHuman GuideOpenHuman Guide

Capabilities

OpenHuman Features

Memory, integrations, local storage, model routing, and native tool access — the five areas that determine whether OpenHuman fits your workflow.

Install guidePrivacy checklist

OpenHuman Features

Practical notes for evaluating a fast-moving open-source AI assistant.

Practical, source-linked OpenHuman guidance

At a Glance

OpenHuman offers five core capabilities that determine whether it fits your workflow: a hierarchical Memory Tree for persistent cross-source memory, 118+ OAuth integrations for app connectivity, TokenJuice for LLM context compression, intelligent model routing between local and cloud AI, and a desktop mascot with voice interaction. Local-first storage is the defining architectural choice — your data lives in SQLite and Markdown on your machine, not in a remote vault.

  • Memory: three-layer hierarchical tree (themes, entities, documents) in local SQLite + Obsidian-compatible Markdown vault.
  • Integrations: 118+ OAuth connectors with background sync every ~20 minutes.
  • TokenJuice: 70-80% reduction in LLM context usage via compression pipeline.
  • Model routing: automatic local/cloud switching based on workload type.
  • Voice: Whisper STT (local), ElevenLabs TTS (cloud), animated desktop mascot with Google Meet participation.

Feature Areas

Memory Tree

A local-first memory tree and Obsidian-compatible Markdown vault built from your connected accounts.

Open

Integrations

118+ connectors via one-click OAuth for email, calendar, code, docs, chat, and more.

Open

TokenJuice

Token compression that shrinks tool output before it hits the model's context window.

Open

Model Routing

One subscription unlocks 30+ AI providers, with automatic routing between local and cloud based on workload type.

Open

Voice & Mascot

Animated desktop mascot with Whisper STT, ElevenLabs TTS, lip sync, and optional Google Meet presence.

Open

Memory Tree

OpenHuman's Memory Tree is a deterministic, bucket-sealed pipeline — not a thin vector-database wrapper. Data from connected accounts converts to Markdown, splits into scored chunks of up to 3,000 tokens, and folds into a hierarchical summary tree stored in local SQLite and an Obsidian-compatible vault.

  • Three conceptual layers: themes (work, family, finance), entities (people, companies, repos), and raw documents (emails, notes, transactions).
  • Relationships between layers are kept explicit so the agent can answer cross-source questions without reloading full context.
  • The vault is inspectable and editable: open it in Obsidian, read or delete chunks, and the agent picks up changes on the next ingest cycle.

Integrations & TokenJuice

118+ third-party integrations via one-click OAuth: Gmail, Calendar, Drive, Notion, GitHub, Slack, Stripe, Linear, Jira, Telegram, Discord, Zoom, Outlook, Dropbox, and more. After the initial exchange, tokens store encrypted in the local vault. TokenJuice then compacts tool output before it reaches any LLM.

  • TokenJuice pipeline: HTML to Markdown, URL shortening, non-ASCII removal, deduplication, and key-info extraction.
  • Maintainer claim: up to 80% reduction in cost and latency. A single-reviewer benchmark reported roughly 70% reduction in one scenario.
  • Background sync runs roughly every 20 minutes for each active connection, folding new content into the Memory Tree.

Model Routing & Local AI

One subscription unlocks 30+ AI providers. The router sends reasoning, agentic, and coding tasks to cloud frontier models; reaction, classification, formatting, sentiment, summarization, and lightweight tasks route locally when local AI is enabled.

  • Local AI is opt-in and off by default. When enabled, supported workloads route to Ollama or LM Studio.
  • Vision tasks are routed to a vision-capable model regardless of local/cloud setting.
  • Optional agentmemory backend lets OpenHuman share the same durable memory store with Claude Code, Cursor, Codex, and OpenCode.

Evaluation Notes

  • Only connect the integrations you actually need — every connector adds attack surface.
  • Know where each task runs: local model, bundled subscription, or external provider.
  • Treat memory as sensitive: it summarizes personal and work data.
  • Voice and Meet use ElevenLabs (cloud TTS) and the Meet API — local-first does not mean local-only.
What are OpenHuman's main features?

Five core capabilities: a three-layer Memory Tree for persistent cross-source memory, 118+ OAuth integrations, TokenJuice context compression, intelligent routing between local and cloud models, and a desktop mascot with voice. Data is stored local-first in SQLite and an Obsidian-compatible Markdown vault.

What is TokenJuice and how much does it save?

TokenJuice is a compression pipeline (HTML to Markdown, URL shortening, non-ASCII removal, deduplication, and key-info extraction) that compacts tool output before it reaches the model. The maintainer claims up to 80% reduction in cost and latency; a single-reviewer benchmark reported roughly 70% in one scenario.

Does OpenHuman run AI models locally?

Local AI is opt-in and off by default. When enabled, lightweight workloads such as reaction, classification, formatting, and summarization route to Ollama or LM Studio, while reasoning, agentic, and coding tasks go to cloud frontier models. Vision tasks always route to a vision-capable model regardless of the setting.

How does OpenHuman's memory work?

Data from connected accounts converts to Markdown, splits into scored chunks of up to 3,000 tokens, and folds into a hierarchical summary tree of themes, entities, and documents in local SQLite and an Obsidian-compatible vault you can open, read, or delete. The agent picks up edits on the next ingest cycle.

Can OpenHuman share memory with coding agents?

Yes. An optional agentmemory backend lets OpenHuman share the same durable memory store with Claude Code, Cursor, Codex, and OpenCode.