Kiro Crew Memory, Schedules, and Heartbeats
On this page
Quick answer
Kiro Crew’s memory, lessons, schedules, webhooks, heartbeats, Task Runner, and subagents turn a chat into a persistent operating system for work. Configure each as a durable automation with a named owner, trigger, scope, authority, review window, stop path, and cleanup plan.
Persistence improves continuity. It can also preserve bad assumptions, repeat an action, carry secrets or sensitive context forward, and keep working after the original reviewer leaves.
Separate the persistence types
| Surface | Persists | Main risk |
|---|---|---|
| Memory | Preferences and relevant context | Stale, sensitive, or cross-project context influences later work |
| Lesson | A correction applied in future sessions | One local exception becomes a broad rule |
| Skill | A reusable procedure | Procedure and permissions drift after dependencies change |
| Schedule | A recurring trigger and task | Repeated cost or action continues without review |
| Webhook | Event-driven task start | Untrusted or replayed events create work |
| Heartbeat | Repeated monitoring until attention is needed | Polling scope or notification freshness is misunderstood |
| Task checkpoint | Progress and retry state | Resume repeats external effects |
Visible and editable memory is valuable, but visibility alone does not make stored context correct or appropriate. Review source, scope, sensitivity, age, and deletion behavior.
Design schedules and webhooks
For each automation record:
- exact trigger, timezone, cadence, and start/end date;
- authenticated webhook source and replay protection;
- allowed repositories, data, tools, commands, networks, and destinations;
- maximum model, infrastructure, and third-party cost;
- approval before any irreversible or external action;
- idempotency key and duplicate-effect check;
- timeout, retry, checkpoint, and escalation behavior;
- owner, reviewer, notifications, and emergency disable path.
Jobs that do not require reasoning can run as plain scripts or commands according to Kiro’s launch. That can reduce model cost, but it does not reduce the need for input validation, least privilege, logs, and recovery.
Heartbeats are monitors, not transactions
A heartbeat can watch a PR, deployment, issue queue, or other condition on a schedule. Define:
- what source is checked and with which credentials;
- what exact state counts as changed or unhealthy;
- how stale data and check failures are distinguished;
- where alerts go and who acknowledges them;
- whether the heartbeat may act or only notify;
- when monitoring stops and how that is verified.
Do not use a non-urgent heartbeat as the only protection for an irreversible action. Monitoring can fail silently, alert delivery can lag, and a stop request may not undo an action already taken.
Control subagents and Task Runner
Independent sessions and subagents should receive only the context and tools needed for their subtask. The parent should validate returned evidence before applying it. Parallel work increases throughput and also expands simultaneous credential, tool, and cost exposure.
Task Runner can decompose, execute, test, retry, and checkpoint longer work. Make external effects idempotent and separate “test passed” from “change approved.” A retry after a timeout must check whether the original action succeeded.
Start with the Kiro Crew overview, verify the Gateway topology, and use the security readiness check before unattended operation.
Frequently asked questions
What persists in Kiro Crew?
Kiro documents persistent sessions, memory, lessons, skills, schedules, and task checkpoints that can survive beyond one chat and across Gateway restarts.
What is a Kiro Crew heartbeat?
A heartbeat is scheduled monitoring that checks a system or task until something needs attention and then alerts through a configured destination.
Does deleting a schedule cancel work already started?
Do not assume so. Verify the trigger is disabled, inspect already-created sessions and external actions, stop active work, and confirm the resulting system state.
Official sources
Source check: August 19, 2026. Recheck memory, lesson, skill, schedule, webhook, heartbeat, Task Runner, subagent, checkpoint, notification, and deletion behavior before relying on automation.