KaziKazi

How do I get an AI agent to remember a correction on its next scheduled run?

Run every occurrence of the job on the same thread, so the next run starts with the last run's history, including your correction. Many tools start each scheduled run in a new chat, which is why you end up correcting the same mistake every Monday. In Kazi a scheduled subagent keeps one thread across all of its runs.

Ken Wiltshire, founder of Kazi · Checked against the product on 2026-11-14

Why corrections disappear

A model doesn't learn between calls. Nothing you say changes its weights. The only way a correction carries forward is if it's in the context the next run reads.

So the question is what the next run reads. If the scheduler opens a new chat each time, the answer is the task description and nothing else. Your correction lived in last week's chat, and last week's chat is gone from view.

What one thread per job looks like

Say Kazi chases unpaid invoices every Monday. The first run emails a client you'd already agreed to give extra time. You tell it, in that thread: "Acme is on 60-day terms, skip them until the 15th."

Next Monday's run starts on the same thread. Your note is in its history, so it skips Acme. You made the correction once.

Long threads would eventually outgrow any context window, so Kazi seals older turns into memory and recalls them when they matter. The thread has no fixed limit.

When to write it down instead

A correction in the history works, but it's implicit. When a rule matters, like "never email a client on a payment plan", put it in the job's playbook too. History is how the agent learns; a playbook is how you make sure.