Agent memory

An agent's model starts each conversation without the last one. Agent memory is the layer that stores what the agent learned and returns the part that matters for the next question. It is more than a copy of the transcript: it knows which value is current and which it replaced.

• Keep each user's memory apart from every other user's.
• Replace a fact when the fact changes, and keep the old value answerable by date.
• Name the source of every answer.

• Ask whether it keeps dates: a store with no event time cannot tell old from new.
• Ask whether a delete reaches the memories derived from the deleted text.
• Ask for the evidence behind an answer, not only the answer.

No. A vector database finds text that looks like a question. Agent memory also tracks which fact is current and when it changed.