The fact changed.
The agent answered anyway.
An agent is told something. Later it is told otherwise. Then someone asks, and the agent answers from the old fact with full confidence. Step through it twice: once with retrieval over what was said, once with Polign Recall.
Run 1: retrieval over everything said
Answers staleThe usual setup. Each message the agent sees is embedded and stored. At answer time it searches for what looks closest to the question and hands the top results to the model.
Run 2: the same steps with Polign Recall
Answers currentThe agent records the fact as a typed statement. is declared single-valued, so the second statement supersedes the first by rule. Nothing is deleted. The read returns only what holds now.
What the model was given
Both runs used the same model and the same question. The only difference is what memory handed it.
Retrieval
Polign Recall
Try it on your own machine with the quickstart, read how corrections work on the Polign Recall page, or bring an agent whose facts change.