Skip to content

Orchestrator/ Memory: Implement review feedback memory loop so agents learn from PR reviews #27

@leandrodamascena

Description

@leandrodamascena

Component

API or orchestration

Describe the feature

This issue depends on #26

This addresses the "Review feedback memory loop (Tier 2)" item from Iteration 3d in the roadmap.

The agent doesn't learn from PR reviews. If a reviewer leaves comments like "use zod for validation", "return 201 for creation endpoints", "use our errorHandler middleware" - those corrections apply to every future task on that repo, but the agent has no way to internalize them. The next task on the same repo will repeat the same mistakes, and the reviewer has to leave the same comments again.

The memory system (Iteration 3b) already supports writing and reading repo-level knowledge. The webhook infrastructure (Iteration 3a) already handles inbound events. What's missing is the pipeline that connects PR review comments to persistent memory.

Use case

I submitted a new_task to create an API endpoint. The agent delivered working code but without request validation, wrong HTTP status codes, and inline error handling instead of the repo's middleware pattern. I left 3 review comments. Then I submitted another task for a different endpoint on the same repo - the agent made the exact same mistakes because it had no memory of my feedback.

Proposed solution

Use the existing webhook infrastructure to capture pull_request_review events and extract actionable rules into memory.

Flow:

  1. Reviewer leaves comments on an agent-created PR
  2. GitHub sends a pull_request_review webhook event
  3. A new handler (separate from the task-creation webhook) receives the event, fetches the diff + comments
  4. A Bedrock call extracts rules from the comments, classifying each as:
    • Repo-level: "use zod for validation on all endpoints" - applies to all future tasks
    • Task-specific: "this function needs a null check" - context for this task only
  5. Rules are written to AgentCore Memory with source_type: review_feedback provenance
  6. On the next task for that repo, loadMemoryContext() loads these rules during hydration - the agent starts already knowing the repo's conventions

A real example:

Review comments: "validate with zod", "return 201 for POST", "use errorHandler middleware"

Next task prompt includes:

Repository knowledge (from past reviews):
- Use zod for request body validation on all endpoints
- POST endpoints that create resources must return HTTP 201
- Use errorHandler middleware, not inline try/catch

The agent applies all three without the reviewer repeating themselves.

What's we need to add:

  • A webhook handler for pull_request_review events (the current handler only creates tasks)
  • An extraction Lambda with a Bedrock prompt to turn review comments into structured rules
  • Provenance tagging on memory writes (source_type: review_feedback) so trust scoring (Iteration 3e) can weight these appropriately

Other information

No response

Acknowledgements

  • I may be able to implement this feature
  • This might be a breaking change

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions