1. An agent is designed to help users draft and iteratively refine long technical documents across multiple sessions. After the first session ends, the agent must remember the overall document structure and user-approved section headings, but does NOT need to retain the raw intermediate drafts the user rejected. Which memory strategy best fits this requirement?
- A. Store the entire conversation history including all rejected drafts in short-term memory so the agent can reference the full context on next launch.
- B. Persist only the approved document structure and section headings as long-term memory artifacts, discarding intermediate drafts after each session.✓ Correct
- C. Use external vector storage to embed every message in the conversation and retrieve all of them at the start of the next session.
- D. Rely on the model's in-context window across sessions by passing the full prior session transcript each time the user reconnects.
Explanation
Option B is correct because persisting only task-relevant, user-approved artifacts (document structure and headings) as long-term memory satisfies the requirement to remember decisions without retaining unnecessary rejected content — this is scoping agent memory to task-relevant information. Option A is wrong because storing all rejected drafts in short-term memory wastes context and violates the principle of scoping memory; short-term memory does not persist across sessions anyway. Option C is wrong because embedding every message — including rejected drafts — in an external store contradicts the goal of discarding non-relevant content and would cause stale, unwanted context to influence future decisions. Option D is wrong because re-injecting the full prior transcript each session bloats the context window and makes the agent process information the user explicitly rejected, risking context drift.