Skip to content

OpenHuman

🔍 Quick overview

OpenHuman is an open-source agent (GPL-3.0) based on Tauri + Rust, shipped as a desktop app, a terminal, or a headless server. As of 2026-10-09, its memory is no longer the local Memory Tree of the early versions: documents, conversations, and “learnings” are stored in CortexDB, either hosted by TinyHumans (login required) or on a CortexDB endpoint you operate; before each turn, a token-bounded “memory pack” is recalled with citations12.

However, it is not a 100% on-premise tool by default. The documentation is explicit: by default, the OpenHuman backend relays LLM calls, OAuth tokens (119 applications via Composio), web search, and hosted memory; even with local models, login, billing, and integration management go through that backend34.

💡 Why this project interests us

OpenHuman was, up to its version 1, the most readable example of a “memory-first” assistant: it structured documents into summary trees and kept a human-readable Markdown equivalent.

For this vault, it served as the reference for the Memory Tree pattern up to its version 1; since memory v2 (CortexDB, 2026), it mainly illustrates two other lessons: an assistant’s memory can leave the machine even when the model is local, and a “local-only” mode enforced in code is worth more than a configuration53.

✅ Strengths

  • Pluggable memory: hosted or self-hosted CortexDB engine, “learnings” and “beliefs” built in the background, recall via memory pack with citations; secrets and credentials scrubbed before storage, full erasure possible2.
  • Local-only Privacy mode: a switch enforced in the Rust core refuses every cloud provider (including BYOK and Composio) and only lets through the local runtimes Ollama, LM Studio, MLX, or OpenAI-compatible5.
  • Tooled agent: search, web fetch, files, Git, lint/test/grep, integrations, and voice per configuration6.
  • Local AI possible: Ollama/LM Studio can take some workloads on-device6.

⚠️ Limitations and risks

  • Cloud by default for several critical functions: LLM routing, proxy web search, managed OAuth/integrations6.
  • Non-trivial sovereignty: managed paths must be replaced one by one.
  • Fast-changing product: still “early beta” according to its authors despite strong adoption (GPL-3.0, near-daily releases, 674 PRs for v0.64.0 in September 2026); a managed subscription (Free, Basic $19.99/month, Pro $199.99/month) funds the hosted model routing7.
  • Broad integration surface: Gmail, Slack, GitHub, Notion, etc. require strict permission governance.

🔒 Sovereignty and privacy

  • Data: workspace files, settings, and audio buffers on the machine; memory, no: it lives in CortexDB (TinyHumans-hosted or an endpoint you operate)3.
  • Model: managed routing by default; Ollama/LM Studio possible for local workloads6.
  • Memory: CortexDB, hosted or self-hosted; without a login or CortexDB key, memory is disabled23.
  • Telemetry: to verify in deployed instance.
  • 100% offline mode: partial; managed features and real-time integrations may require the backend.
  • Verdict: ⚠️ configurable.

See the full grid: Sovereignty & Privacy.

🔗 Possible integration in this vault

OpenHuman is relevant as:

  • a reminder that an assistant’s memory can leave the machine even with a local model;
  • a comparison page to explain the “local-first ≠ sovereign by default” trap;
  • an example of a hybrid solution not to present as strict on-prem without caveat.

📊 Project maturity

Open-source project (GPL-3.0) evolving very rapidly (v0.57.5 in June 2026, v0.64.15 as of 2026-10-09), repositioned as an agent harness (desktop, terminal, headless server, Rust library). Audit before client deployment: OAuth token retention on the backend side, dependency on the backend for login, billing, and integrations even in local-only mode, and self-host options (your own CortexDB, headless core in Docker)37.

🔗 See also

📚 Sources

Footnotes

  1. OpenHuman, Architecture — React + Tauri v2 desktop app, Rust core running in the same process, standalone terminal frontend (openhuman-tui), openhuman-core binary and embeddable library (openhuman-embed) (read on 2026-11-08). https://tinyhumans.gitbook.io/openhuman/developing/architecture ↩

  2. OpenHuman, How memory works — “The current memory has no memory tree … or Obsidian vault”; TinyHumans Hosted / CortexDB engines, memory pack, secret scrubbing (docs read on 2026-10-09). https://tinyhumans.gitbook.io/openhuman/features/memory ↩ ↩2 ↩3

  3. OpenHuman, Privacy and security (where memory lives, what the backend relays), docs read on 2026-10-09. https://tinyhumans.gitbook.io/openhuman/features/privacy-and-security ↩ ↩2 ↩3 ↩4 ↩5

  4. OpenHuman, Local and BYOK models (model routing: login, billing, and integrations via the backend even with local models), docs read on 2026-10-09. https://tinyhumans.gitbook.io/openhuman/features/model-routing/local-and-byok-models ↩

  5. OpenHuman, Privacy mode (local_only, enforced in the Rust core; allowed local runtimes), docs read on 2026-10-09. https://tinyhumans.gitbook.io/openhuman/features/privacy-mode ↩ ↩2

  6. OpenHuman README — “Local + managed services, upfront”, Ollama/LM Studio, Composio, and managed backend. https://github.com/tinyhumansai/openhuman/blob/main/README.md ↩ ↩2 ↩3 ↩4

  7. OpenHuman, Releases (v0.64.15 as of 2026-10-09, 674 PRs for v0.64.0 “Intelligence Upgrade” of 2026-09-25) and Billing and usage (Free, Basic $19.99/month, Pro $199.99/month plans), read on 2026-10-09. https://github.com/tinyhumansai/openhuman/releases · https://tinyhumans.gitbook.io/openhuman/features/billing-and-usage ↩ ↩2

Give my opinion

Everything is optional: one answer is enough. This site sends nothing: you send it, the way you choose.

Your noteNo note
Your opinion is about
About what? (several choices possible)
What will be sent
Send by e-mail

Private, no account: opens your mail program with the message ready, for ia-on-prem@damien.becherini.fr.

Open a GitHub issue ↗

Public, with a GitHub account. You read the issue again before you publish it.

To paste it wherever you want.

Give at least a note, a subject or a message to be able to send.