Skip to content

[Feature Request] Memory transfer between agent instances — share learned patterns across agents #211

Description

@BabjiRishi

Problem

Azure SRE Agent memory (learned patterns, investigation findings, operational context) is agent-local — it never transfers between agent instances. When promoting agent configurations from development to test and production tiers, only repo-based artifacts (skills, knowledge files, custom agents, tools, triggers) are portable. The valuable patterns the agent learned from hundreds of investigations stay locked in the source agent.

Current Behavior

  • What promotes via CI/CD: Knowledge files, skill files, custom agent definitions, tool definitions, trigger configs, MCP connector configs, scripts — all sourced from a git repo
  • What does NOT promote: Conversation threads, agent memory (learned patterns from investigations), session insights — all are agent-local and geography-local
  • Workaround: Manually export valuable findings as .md files → commit to a shared repo → CI/CD deploys to all agents. This is labor-intensive and loses the semantic richness of the agent's learned context.

Proposed Solution

  1. Memory export/import API — programmatically export an agent's learned memory as a portable artifact (e.g., structured JSON or markdown bundle) and import it into another agent instance
  2. Selective memory sync — allow administrators to sync specific memory categories (e.g., "service architecture", "incident patterns", "team preferences") between designated agents
  3. Memory promotion in CI/CD — integrate memory export into CI/CD pipelines so that validated patterns from development agents automatically promote to production agents alongside skills and knowledge

Use Case

Multi-Tier Deployment

Organizations operating tiered agent deployments (dev → test → prod):

  • Dev agents accumulate deep operational knowledge through daily investigations
  • When prod agents are provisioned, they start from zero learned context — they don't benefit from months of dev investigation history
  • Engineers must re-teach the same patterns, team preferences, and service-specific context to each new agent

Multi-Agent Topology

With multiple product-specific agents, investigation patterns discovered by one agent (e.g., "database connection pool exhaustion correlates with a scheduled job") are valuable to other agents but remain siloed.

Impact

  • Medium — accelerates new agent onboarding from days to minutes
  • Enables knowledge compounding across the entire agent fleet
  • Critical for production readiness: prod agents should inherit dev agents' operational maturity

Environment

  • Azure SRE Agent (GA)

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type
    No fields configured for issues without a type.

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions