Github Copilot's Agentic Memory System

intro

Vision: Enable Github Copilot to become an ecosystem of agents that collaborate.
Multi-agent workflows - moving beyond isolated interactions and toward a cumulative knowledge base that grows with every use.
Cross-agent memory: agents remembering and learning from experiences across the development workflow without relying on user instruction.

Challenge: what to remember and when to forget.
Memory system must handle code changes, abandoned branches, and conflicting observations.
One could have an offline curation service, but this is complex and expensive at large scale.
solution: Just-In-Time (JIT) verification
information retrieval is hard to solve but easy to verify
store memories with citations, references to specific code locations. When an agent encounters a memory, it verifies the citation in real-time before using it
memory creation as a MCP tool call - Agents invoke the tool when they discover something likely to have implications for the future.
Agent can call a store_memory tool, which invokes a memory API and writes to the memory DB
example - While reviewing a PR, Copilot discovers that API version must be synced across three different parts of the codebase. It can write a memory to remember this information

at the start of a session, retrieve the most recent memories for the target repo and include it in the prompt
before applying a memory, the agent verifies accuracy and relevance by checking citations. When there is a contradiction, the agent stores the corrected version with new evidence
for security, memories stay within the targeted repo and are only created through actions by contributors with permission
biggest concern: impact of outdated or bad memories
- found that agents consistently verified citations, and the memory pool would self-heal