Agent knowledge loops

When to Use

Load this when the task is "make each new agent better than the last one", "write down what the agents learn", "stop them rediscovering the same things", "accumulate conventions across repos", or when you are designing any feedback path from task outcome into what the next run reads. Not for one-off prompt tuning, and not for editing a single conventions file by hand.

Two artifacts, never one

The most common design error is conflating them. When the user says "ledger", ask whether they mean the bookkeeping or the knowledge; they are different artifacts and usually live in different places.