Memorize[Examples]

A memory of approved and rejected agent outputs paired with the input context, the reasoning, and why a human accepted or rejected the work. New tasks retrieve the closest real precedent instead of improvising from a synthetic few-shot. Surfaced to agents over MCP. Useful where the right answer is subjective: writing a sales email, deciding to send a notification, classifying lead quality, handling a sensitive reply.

Why it matters

For subjective work (writing a sales email, deciding whether to send a notification, judging lead quality, handling a sensitive reply) a generic few-shot prompt improvises from nothing. Memorize[Examples] gives agents real precedent: the actual outputs your team approved or rejected, and why. The work drifts toward your standards over time instead of restarting from a synthetic example every run.

How it works

It stores approved and rejected agent outputs paired with the input context, the reasoning, and the human's accept or reject decision. A new task retrieves the closest real precedent rather than a hand-written few-shot, surfaced to agents over MCP. The judgment that used to live only in a reviewer's head becomes retrievable memory.