Operators notice a long-running multi-agent investigation “forgets” earlier entities: names morph, amounts change, and later agents contradict earlier tool results after repeated summarization. Which diagnosis and response is most appropriate?

Diagnosis is summary drift and entity continuity failures from aggressive context compaction. Multi-agent investigations can suffer sliding-window amnesia, summary drift, vector-only recall limits, and entity continuity problems. Adjust compaction prompts, preserve structured entity state outside free-text summaries, and add evaluations that detect entity/value inconsistencies across steps. Increasing temperature so the model invents more entities is wrong because it raises variability, not factual continuity. Deleting structured state stores and keeping only compressed prose summaries is wrong because prose-only memory worsens drift. Assuming Azure region latency is the sole cause is wrong because latency does not explain mutating entities across summaries. Monitor diffs of key entities between workflow stages and fail builds when continuity checks regress.