Memory

Every incident becomes a test, every promise a production check, and AI agents ship only where they have earned it.

Open the Memory Console

Neveragain.

Judge Mode replaying incident #2: the immune loop ring with its six stages, requests failing at 40.1% with pool 5, and the gate to production.
40.1%of orders failed when the pool was 5. Incident #2, fixed by an agent in 2 min 34 s.

Fixed in three minutes

74 seconds
80 more seconds
About 7 minutes
About 9 minutes

First the Incident Responder reads the incident, checks past ones and names the cause.

Then it writes the fix with a regression test.

Then the pipeline replays past incidents on a review revision, checks every promise and weighs the agent's earned trust.

Last, the canary sends 10% of production to the new version, then promotes it or rolls it back, and the agent's trust moves with the result.

What it runs on

The promise from issue #3: orders above $500 wait for verification. Check it against this service now.

Or: curl -X POST https://orders-api-rjpdzrsxra-uc.a.run.app/orders -H 'content-type: application/json' -d '{"item":"laptop","quantity":1,"total":650}'
8

GitLab Duo Agent Platform

  • Incident Responder
  • Close the Loop
  • Contract Keeper
  • Contract Breach
  • Release Council
  • Memory Oracle
  • Code Review flow
  • Fix CI/CD Pipeline flow
9

GitLab

  • Merge request pipelines
  • SAST and IaC scanning
  • Secret detection
  • Dependency scanning
  • Container scanning
  • Review revisions
  • Environments
  • Releases
  • Approvals and labels
7

Google Cloud

  • Cloud Run
  • Cloud SQL
  • Artifact Registry
  • Secret Manager
  • Cloud Monitoring
  • Cloud Scheduler
  • Workload Identity Federation
6

The memory

  • Immune replay
  • Living contracts
  • Trust ledger
  • Canary analysis
  • Carbon per pipeline
  • Read-only MCP server
The 3D memory core: incident #2 on the outer ring with five of six stages done, and rings for immune memory, promises and trust inside. The Memory Console: is it safe to let the agents ship, with the trigger DB_POOL_SIZE=5, the cost of 40.1% failed orders and 0 of 40 after the fix. The Memory Console on a phone.

A postmortem should start a memory.

Here, every incident becomes a test that runs before each change, and every promise a check that runs against production. Production runs revision orders-api-00089-kix, with a database pool of 20 per instance. When MR !1 is approved, that number changes through the canary, and every later change to it is replayed against incident #2.

Open the Memory Console
20connections per instance today