Users report that a long-running multi-agent project assistant forgets earlier constraints after many turns. Engineers notice summaries gradually omit the original budget cap while vector recall returns loosely related notes. Which diagnosis and remediation pairing is best?
Select an answer to reveal the explanation.
Short Explanation
This is classic long-context pathology: summary drift drops the budget cap, and vector-only recall is fuzzy on hard constraints. Keep critical entities in structured state, fix compaction to preserve them, and stop relying on vectors alone. Option A. GPU heat (B), deleting memory (C), and fonts/DNS (D) miss the actual failure mode.
Full Explanation
Optimizing multi-agent systems includes diagnosing context window issues such as sliding-window amnesia, summary drift, vector-only recall limits, and entity continuity problems. Option A correctly names summary drift/entity continuity and applies structured memory plus better compaction.
Options B–D are unrelated distractors that do not address context management.
Practical fix: maintain a canonical project state object (budget, deadlines, owners) updated transactionally; summaries must merge into that object rather than replace it silently.
Exam tip: Soft summaries are lossy — pin mission-critical constraints outside free-text compression.