Memory for AI operations and back-office agents

For agents calling internal APIs: Retrack turns tool errors into procedures, stops retry loops and turns lessons into CI tests. Tool errors −67%.

Updated

Operations and back-office agents call internal APIs all day: update records, cancel orders, issue refunds, reroute shipments, roll back deploys. Their APIs have unwritten rules (uppercase reason codes, amounts in paise rather than rupees, a PIN code in its own field, a 409 that means "already shipped") and every rule costs a failed call the first time.

With Retrack, it costs one failed call once.

What Retrack does for an ops agent

  • Turns tool errors into procedure. An error the agent recovered from becomes a pitfall in the playbook; an extra step a task needed becomes a conditional step.
  • Stops retry loops. memory.protect(tools) returns the known fix for a known failure and stops a call that cannot succeed, instead of letting the agent retry until its budget runs out.
  • Understands general-purpose tools. Agents built around a shell, SQL or HTTP tool get memory per command (kubectl rollout undo, update orders, POST /v1/orders/{}/refund), not one blurred memory for "the shell".
  • Guards against regressions. Every playbook and recurring failure becomes a test for retrack test in CI, and retrack migrate checks your procedures on a new model before you switch.

Measured on an ops agent

On the Ops Desk benchmark, a Claude Haiku 5.5 operations agent working against APIs with unwritten rules, Retrack cut tool errors by 67% and tool calls by 17% across 20 paired tasks, with memory recalled on 19 of them. See the full results.

Good outcome signals for ops

Ops tasks usually have a checkable end state, the best signal there is:

  • the record, order or deployment is in the expected state,
  • a downstream job succeeded,
  • an assertion in your own code passed (source="test", the strongest signal).

Trust your internal tools

Internal APIs whose output you control can be marked as trusted in your project's settings. Runs that use only trusted tools, with a hard success signal, can promote playbooks on their own after three confirmations, with no manual approval. See the memory firewall.

Start with the quickstart.