Yufeng Li and colleagues test what happens when an AI assistant keeps saving new versions of the same old claim. Those copies can occupy several retrieval slots and outcompete a correction. Their Memory Echo study separates remembering a claim, retrieving it and actually using it in an answer.
The proposed fix tracks where each record came from:
- Serve one representative per original source, so paraphrases gain no extra votes.
- Keep superseded records for audit, but exclude them from current answer evidence.
- Supply verified replacement evidence and check the answer against it.
Across 190 source questions and four memory systems, the authors report accuracy rising from 82.5% to 92.0% with DeepSeek V4 Flash. Kimi K2.6 rises from 78.2% to 91.6%. Each comparison replays 1,520 paired states derived from those questions.
Filtering alone can hurt. With missing or unreliable corrections, accuracy falls from 37.8% to 21.2%; travel planning also gets worse when useful history is withheld. A memory system needs both a reason to retire an old claim and usable evidence to replace it.
Caveat: these are controlled histories with supplied source labels and model-based grading; part of the gain comes from selecting verified answer text. The public code excludes the datasets and result artifacts.