GitHub Copilot Gains Cross-Agent Memory System in Public Preview

Alvin Lang
Jan 15, 2026 22:07

GitHub launches memory feature for Copilot agents, enabling AI assistants to learn from past interactions and share knowledge across coding, CLI, and code review workflows.

GitHub has rolled out a cross-agent memory system for Copilot that allows its AI assistants to retain and share knowledge across development workflows. The feature, now in public preview for Copilot Pro and Pro+ subscribers, marks a significant shift from isolated AI sessions toward persistent, cumulative learning.

The memory system works across three Copilot agents: coding agent, CLI, and code review. When one agent learns something about a codebase—say, how database connections are handled—other agents can access that knowledge for future tasks.

How It Actually Works

Rather than building complex offline curation systems to manage stored memories, GitHub’s engineers opted for real-time verification. Each memory gets stored with citations—specific code locations that support the information. When an agent retrieves a memory, it checks those citations against the current codebase before acting on it.

“Information retrieval is an asymmetrical problem: It’s hard to solve, but easy to verify,” wrote Tiferet Gazit, principal ML engineer at GitHub, in the technical blog post announcing the feature.

This approach handles a thorny problem: code changes constantly. A logging convention observed in one branch might get modified, superseded, or never merged at all. By verifying citations in real-time, agents avoid acting on stale information.

Early Performance Numbers

GitHub ran A/B tests measuring the impact on real developer workflows. The results showed a 7% increase in pull request merge rates for Copilot coding agent (90% with memories versus 83% without) and a 2% bump in positive feedback for code review comments. Both improvements registered as statistically significant with p-values below 0.00001.

The company also stress-tested the system by deliberately seeding repositories with adversarial memories containing false information and broken citations. Agents consistently caught the contradictions and updated incorrect memories rather than propagating bad information.

Privacy Boundaries

Memories stay tightly scoped to individual repositories. Only contributors with write permissions can create memories through their actions, and only users with read access can trigger their use. The feature ships as opt-in and disabled by default.

The announcement comes amid other Copilot updates this week, including enhanced CLI agents and context management features announced January 14, plus new bring-your-own-key capabilities. GitHub also disclosed upcoming deprecations for select Claude, Google, and OpenAI models within Copilot.

For development teams evaluating AI coding tools, the memory system addresses a genuine friction point: the repetitive context-setting that wastes time at the start of each AI interaction. Whether the 7% improvement in merge rates holds up at scale remains to be seen as the feature moves beyond preview.

Image source: Shutterstock