Skip to content

Feedback memory

Every decided review is a data point. Feedback memory makes those data points queryable: your agent can read past human decisions before acting, without any additional configuration beyond a standard API call.

GET /api/v1/feedback returns decided reviews in a shape optimized for agent learning loops. Each record carries the original submission, the human’s response, and whether the human changed anything.

Terminal window
curl "https://api.gatewerk.com/api/v1/feedback?template=email-review&outcome=edited&limit=20" \
-H "Authorization: Bearer $GATEWERK_API_KEY"

What query parameters does the endpoint accept?

Section titled “What query parameters does the endpoint accept?”
Parameter Type Default Description
template string (none) Filter by template slug
outcome string (none) Filter by decision value (approved, rejected, edited, etc.)
limit integer (1–200) 50 Number of records to return
offset integer 0 Pagination offset

The response uses the standard { items, total, has_more } envelope.

Each feedback record includes the same triple described in Decisions and webhooks:

  • suggested_value: the payload as the agent originally submitted it
  • approved_value: the payload the reviewer accepted (original merged with any edits)
  • was_edited: true if the reviewer changed any field, false if approved exactly as submitted

Comparing suggested_value and approved_value when was_edited: true gives the reviewer’s precise correction.

Lapsed monitoring windows are excluded. A monitoring review that closed because the veto window elapsed without action is attributed to system:monitoring_window with lapsed: true. These records do not appear in feedback queries. The intent: absence of objection is not a human correction and should not be treated as a training signal.

Notes are free-text sticky context attached to individual reviews, templates, or chain runs. Feedback memory is the structured query surface for decided reviews. The two are complementary: a reviewer might leave a note explaining their reasoning, and that note appears on the review detail; the feedback endpoint returns the input/output delta across many reviews for pattern analysis.

This endpoint does not perform ranking, similarity matching, or any analysis of content. It returns the records that match the filter parameters.


See also: Decisions and webhooks: the feedback triple (suggested_value, approved_value, was_edited) in the webhook payload. Monitoring: why lapsed reviews are excluded from feedback. Notes: per-review context notes.