n8n launches Nod personal-agent preview with 93.2% memory score

n8n's GitHub repository reports a 93.2% LongMemEval score for Nod, up from 79%, using 500 questions about earlier chats answered in a fresh session.

n8n released Nod on November 10, 2026, as a free, open-source research preview for Mac and Windows. The agent runs from a private sandbox on the user's computer and acts in that user's apps and websites. For employers, account delegation, spending approval and data handling become immediate HR and IT policy questions.

What does the research preview change?

Users need no n8n account and no initial server or cloud-service setup, although they must connect an AI provider. n8n published the code under Apache 2.0. The project belongs to n8n Labs, a new team for experiments outside the core product.

Nod is neither an n8n workflow nor the n8n AI Assistant. Existing users can optionally connect an n8n instance so the agent can work with their workflows. n8n said development of the core product and the n8n agent would continue as planned.

What can Nod do without an n8n workflow?

Nod broadens n8n's automation focus from shared team processes to jobs handled by one person. It connects to accounts for business or household work. Nod has its own browser and can sign in to websites without a formal integration. Documented tasks include:

• Cancel a subscription and request a refund.
• Retrieve invoices from supplier portals.
• Watch prices and extract figures from dashboards without export functions.
• Conduct web and PDF research, then create documents, spreadsheets and code in a Linux workspace.
• Read or draft in Gmail, Calendar, Slack, Notion and Linear.

Users can hold one conversation through the app, Telegram or WhatsApp. Several jobs can run in parallel at the same time. Recurring work, including morning briefs and price checks, can run on schedules. After reading connected apps, Nod can also propose jobs.

How does Nod remember past work?

Nod retains months of conversations and records how it completes tasks. It reuses those notes and revises them when a site changes. It can learn a user's writing style from sent Gmail messages. Users can inspect, edit or delete stored information.

LongMemEval tests information extraction, multi-session reasoning, knowledge updates, temporal reasoning and abstention when information is unknown. The benchmark therefore concerns long-term memory, not how reliably Nod completes a booking or refund.

What stops Nod from acting without permission?

Before Nod sends, pays, posts or deletes anything outside the user's request, a separate model-based check asks for confirmation. A webpage or email cannot grant that authority. n8n cautions that the checking model can make mistakes, so the approval step provides a safeguard rather than a guarantee of correct judgment.

Nod encrypts saved logins and payment cards on the machine. It enters them through private forms that the model cannot read. It also logs sign-ins, card use, actions in connected apps and model costs. Those records create an activity trail alongside the approval controls.

Where does employee data go?

Nod keeps its database, files, memory and saved logins on the machine that runs it. The selected AI provider receives the information needed for each task. Local control therefore does not mean every piece of task data stays on the device.

App connections pass through a connection service. Beyond update checks, information reaches n8n only when users enable usage data or submit feedback. Employers will need to distinguish local execution from provider processing and n8n telemetry when setting data-handling expectations.

What will running Nod cost?

Nod itself is free. Users pay their chosen model provider for usage, or they can connect an open-source model running locally. Supported connections include OpenAI, Anthropic, OpenRouter, Amazon Bedrock, Google Vertex AI and Azure OpenAI; ChatGPT Plus and Pro plans are also supported.

n8n lists these cost controls:

• A receipt for each finished task.
• A monthly spending limit.
• An adjustable check-in once a task passes $5.
• An estimated monthly cost for each schedule.

Those controls matter because scheduled and parallel work draws on the user's own model account. The preview does not insulate users from provider charges.

The desktop download requires macOS 14 or later. n8n labels Windows 11 support a preview. Users can also self-host Nod on a personal server. A laptop installation stops when the laptop sleeps. Scheduled jobs therefore need the machine to remain on, which limits unattended use. One installation currently has one owner, and Nod cannot make phone calls.

What should employers test before adoption?

HR and IT teams should test the boundary between assistance and delegation. They need to know which employee accounts may connect, who can approve an action, and whether activity records meet internal oversight needs. They should also test whether employees understand what the agent may do with stored logins, cards and conversation history.

n8n lists a hosted version and shared team use as future work. Until then, local control and provider choice come with direct model costs, an always-on machine for unattended schedules and an approval mechanism that can still make mistakes. The next adoption test is whether those controls can govern real work without leaving every scheduled job dependent on an awake laptop.