GitHub Copilot for JetBrains Gets Persistent Memory and Ollama Support

GitHub has shipped a notable update to its Copilot plugin for JetBrains IDEs, adding persistent memory that carries context between agent chat sessions, Ollama support as a local bring-your-own-model provider, and a new layer of enterprise admin controls. The release also expands Codex workflow visibility, simplifies Copilot CLI installation from within the IDE, and fixes a range of reliability issues across MCP servers, terminals, and cloud agents.

What happened

Feature	Detail
Copilot memory	Retains context across agent chat sessions; toggled in the Copilot settings portal
Ollama integration	Now supported as a BYOK provider with model selection throughout JetBrains
Enterprise controls	Server-based settings for plugin availability, MCP access, permissions, and OpenTelemetry
Copilot CLI setup	Auto-installs from integrated terminals on macOS, Linux, and Windows
Codex debug logs	Codex sessions now visible in agent debug logs; supports updated permission modes

Persistent memory across chat sessions

Previously, every new Copilot agent chat session started fresh. You had to re-explain your project structure, coding preferences, or team conventions each time. The new memory feature stores and recalls that information automatically, so you can pick up where you left off without the repetitive preamble. You can turn it on or off with the Copilot Memory toggle in the settings portal.

Ollama as a local model provider

Ollama is an open-source tool for running large language models locally on your own hardware. GitHub has added it as a BYOK (bring your own key/model) provider, meaning you can now point Copilot at a local Ollama instance instead of, or alongside, GitHub’s hosted models. Model selection is available throughout the JetBrains interface. This matters most for teams with data residency requirements or those who want to test smaller, domain-specific models without sending code to an external API.

More enterprise admin controls

Admins now have server-side controls for managing Copilot across an organisation. The new settings cover which plugins are available, whether users can connect to MCP servers (model context protocol servers, which expose external tools and data to the AI), permission bypass behaviour, and OpenTelemetry observability settings. These give IT and security teams more visibility and lock-down options without requiring per-user configuration.

Codex, CLI, and UX changes

Codex agent sessions now appear in the debug log, and the feature supports custom instructions and skills for adapting agent behaviour per project. On the setup side, the plugin can now auto-install Copilot CLI directly from the integrated terminal on all three major operating systems, cutting out a manual download step.

Smaller UX fixes include restored file and folder references in chat inputs (for Copilot, Claude, and Codex modes), a cap on long model names in the model picker, and a repositioned customisation button at the top of the chat panel. User-facing strings have also been standardised to “Copilot” rather than “Copilot CLI.”

Why it matters

The memory feature alone removes a genuine friction point for developers who use agent workflows daily. Repeating context is not just annoying; it burns tokens and slows down iterative work. Persistent memory brings Copilot closer to how a human collaborator actually operates.

The Ollama integration is a meaningful signal for enterprise buyers. Keeping model inference on-premise is a hard requirement in regulated industries, and GitHub is clearly trying to accommodate that without forcing teams to abandon the Copilot interface they already use. Pair that with the expanded admin controls and this update looks more like a serious enterprise pitch than a feature dump.

For teams already exploring agentic coding, the Codex debug log visibility is worth noting. Debugging agent sessions has been opaque; seeing what Codex actually ran and with what permissions is a basic requirement before trusting it with anything consequential. We covered the broader shift toward agentic coding workflows earlier this year, and tooling like this is what makes that shift practical rather than theoretical.

Our take

Memory in AI chat tools sounds trivial until you have used it. The productivity gain from not re-explaining a monorepo layout or a preferred testing pattern every session is real and measurable. GitHub has implemented it with a clear toggle, which is the right call: memory can introduce privacy concerns in shared or sensitive codebases, and teams need the option to disable it.

The Ollama support is smart positioning. Local models are not going to replace hosted ones for most tasks, but having the option matters to enterprise procurement conversations. If you are helping a client evaluate Copilot for a regulated environment, this is now a credible answer to the “does our code leave our network?” question for AI completion tasks.

If your team uses our AI integration services, it is worth revisiting your JetBrains Copilot setup to take advantage of the new admin controls, especially MCP server configuration, before those settings get pushed organisation-wide by default.

What to do about it

1. Update the GitHub Copilot plugin in your JetBrains IDE to the latest version.
2. Navigate to Copilot settings and enable the Copilot Memory toggle if persistent context would benefit your team.
3. If you run Ollama locally, add it as a BYOK provider in the model settings and test it against your typical prompts.
4. If you are an admin, review the new enterprise managed settings, particularly MCP server access and permission bypass behaviour, before they inherit defaults.
5. Check the agent debug logs the next time you run a Codex session to understand what the agent is actually doing under the hood.

Update the plugin first; everything else follows from there.

Source: GitHub Changelog