GitHub shipped two things to Copilot for JetBrains on 11 August, and they are interesting mostly because of the tension between them.
Memory makes the agent stateful: it retains useful context across chat sessions so you stop re-explaining your project. Ollama support makes the model local: inference on your own hardware, nothing sent to a vendor. One accumulates knowledge about you; the other exists so knowledge about you does not travel.
Copilot memory
The mechanic is simple: Copilot memory can retain and recall useful information across agent chat sessions, so project details and preferences you explained once do not need explaining again. It is toggled with the Copilot Memory option in the Copilot settings portal.
The changelog does not state where memories are stored, how long they persist, or how they are scoped — to a repository, a user, or an organization. Those are exactly the questions a security review will ask, so treat them as unanswered rather than assuming the reassuring answer.
It is worth knowing what memory changes about failure. A stateless agent that misunderstands your project misunderstands it once. A stateful one writes the misunderstanding down and applies it to everything afterwards, and you will not see the moment it was recorded. If your agent starts confidently doing something wrong across unrelated tasks, memory is now on the list of things to check.
Ollama as a BYOK provider
Ollama is now a bring-your-own-key provider in Copilot for JetBrains, with provider configuration and model selection available throughout the JetBrains experience. In practice that means the model can run on your own machine.
This is the more consequential of the two for a particular kind of team: the one whose legal or compliance position has, until now, ruled out AI assistance in the IDE entirely. "The code never leaves the building" is a policy that local inference can actually satisfy, and it is the only configuration that satisfies it.
Set expectations honestly. A model small enough to run on a laptop is not a frontier model, and the gap shows most on exactly the work agents are best at — multi-file reasoning and long-context tasks. Local inference is a compliance answer first and a performance answer never.
What the entry does not say
GitHub's post names neither the specific JetBrains IDEs nor the plans these features reach, and it does not describe the default state for a new user beyond the settings toggle. It also refers to enterprise managed settings without spelling out which controls apply.
If you administer a Business or Enterprise plan, do not assume memory is off by default because you would prefer it to be. Check the settings portal yourself, and check it before someone asks you in a review.
The pattern worth noticing
JetBrains users have spent most of the agent era a step behind VS Code. This entry is part of a visible catch-up — and the shape of it is telling: not a new model, but statefulness and provider choice, which are the two things that decide whether a tool can be adopted at all in a regulated team.
The capability race has largely moved to configuration. What you can turn off is becoming as important a feature as what you can turn on.