Skip to content
NewsProduct launch

n8n 2.39.0 ships Instance AI memory, sub-agent sandbox sharing, and a source control API

· by Pondero Newsdesk

The short version

n8n released version 2.39.0 on September 8 as a pre-release, adding persistent conversation memory to Instance AI, workspace folder exploration, background sub-agent sandbox sharing, and public API access to source control push and pull operations.

n8n 2.39.0 ships Instance AI memory, sub-agent sandbox sharing, and a source control API

n8n's built-in AI assistant lost all session context the moment an editor tab closed. Version 2.39.0, released September 8 as a pre-release, changes that: Instance AI can now search its own conversation history and browse workspace folders. The same release lets background sub-agents share their parent agent's execution sandbox and exposes source control push and pull to the public API.

What changed in 2.39.0

Instance AI gained two context capabilities per the 2.39.0 release notes. Conversation search lets the assistant retrieve prior sessions when working on a workflow it has seen before, rather than starting cold each time. Folder exploration extends the assistant's awareness to the full workspace directory, not only the workflow currently open in the editor. A new opt-in concurrency cap controls how many Instance AI runs can execute simultaneously on a single instance, and a reporting module adds analytics coverage for AI activity.

Sub-agent execution also changed. Background sub-agents now inherit the parent agent's workspace sandbox. Previously, each sub-agent ran in an isolated context; the shared sandbox lets parent and child agents operate on the same workspace state without passing data explicitly through workflow nodes.

The public API grew three source control endpoints: push a project to Git, pull from Git, and retrieve current source control status. A separate UI control for selecting which workflow changes to promote connects the editor to the same operations.

Two Anthropic-specific reliability fixes land in this release. Agent threads can no longer enter a permanent broken state after a single run error, and errors now remain visible in the editor after a failure. The Anthropic Chat Model node also correctly sends disabled thinking to the provider and parses content-block output. Azure OpenAI classic and Foundry credential configuration for agents received a separate fix.

Three integrations are new or expanded. Jira and Confluence both gain Atlassian Service Account (2LO) authentication, which uses server-to-server tokens rather than user-delegated OAuth. A Databricks Chat Model sub-node ships as a hidden (feature-flagged) option for LLM calls inside Databricks workflows. The LDAP node now escapes expression output in the Custom Filter field, closing a potential injection path.

The Qdrant client was updated and undici v6 dependencies removed, restoring AI request compatibility with Node 26 runtimes. The editor dropped the "Ask AI" tab from the Code node, consolidating AI assistance through Instance AI instead.

Why it matters

Persistent conversation memory in Instance AI closes the most common friction point for teams using the assistant iteratively. Every new session previously required re-explaining the workflow's credential layout, its purpose, and its failure history. Conversation search means the assistant can retrieve that prior context from its own history rather than requiring the user to repeat it.

The sub-agent sandbox change matters for any workflow that chains agents. A parent agent can now set up workspace state once; background sub-agents inherit that state without additional data-passing nodes between them.

Source control via public API means external CI pipelines can now trigger push and pull without opening the n8n editor, removing the manual step that broke GitOps workflows using n8n.

Operators running Anthropic-backed agents should treat the thread-stability fix as a prompt to update. A broken thread previously required manual intervention; the fix makes recovery automatic on the next run. Operators using the LDAP node should also update: expression output in Custom Filter fields was not escaped before this release.

What to watch

The Databricks Chat Model sub-node is feature-flagged in 2.39.0. When it reaches general availability it will give Databricks-hosted models a native path into n8n agent workflows without a generic HTTP Request node.

Sources