Stacks/Coding agent

GitHub Copilot's cross-agent memory

GitHub's Copilot memory system is a production, opt-in knowledge layer shared by the coding agent, CLI, and code review. Agents store facts with code citations and verify those citations at read time instead of running a heavy offline curator. A/B tests on real developer traffic showed a 7% increase in coding-agent PR merge rates and a smaller lift in code-review feedback.

Use case

A shared memory system so GitHub Copilot agents — starting with the coding agent, CLI, and code review — can remember repository knowledge across sessions and learn from one another

Category

Coding agent

Team size

not stated

Orchestrator / framework

Copilot is described as an ecosystem of agents across coding, code review, CLI, and (in the vision) security, debugging, and deployment. Memory creation is a tool call (`store_memory`) that an agent can invoke when it discovers something likely to matter later. At the start of a session, recent repository memories are included in the prompt; the agent is prompted to verify citations before using a memory.

Memory

Cross-agent, repository-scoped memory. Each memory stores a fact plus citations to specific code locations. Before use, the agent verifies those citations against the current branch in real time (the authors say this is a small number of reads and added no significant latency in testing). Incorrect memories are corrected and re-stored; valid ones can be re-stored to refresh their timestamp. Memories are opt-in, off by default, and scoped to the repository: only write-permission contributors can create them, and only users with read permission can use them.

Evals

Agents were stress-tested with adversarial memories that contradicted the codebase; they consistently verified citations and self-healed the pool. In a simulated noisy memory pool, Copilot code review gained a 3% increase in precision and 4% in recall. A/B tests on real developers: coding agent pull-request merge rates went from 83% to 90% (7% increase); code-review positive feedback on comments went from 75% to 77% (2% increase); both results reported as highly statistically significant (p < 0.00001).

Observability

not stated

Monthly cost

not stated

Models

not stated

Observability tools

not stated

Other tools

store_memory toolCitation-based just-in-time verification

Source

https://github.blog/ai-and-ml/github-copilot/building-an-agentic-memory-system-for-github-copilot/

Tiferet Gazit · Jan 14, 2026