# 研究命名"Ghost Memory"：AI智能体重复已过时事实的问题

- 来源：elvis (@omarsar0)
- 发布时间：2026-07-06 21:20
- AIHOT 分数：27
- AIHOT 链接：https://aihot.virxact.com/items/cmr99mkpy028islsmprj2k2lx
- 原文链接：https://x.com/omarsar0/status/2074121191846261022

## AI 摘要

长期运行的AI智能体会重复已过时的事实，新研究将此失效命名为"ghost memory"。旧事实、当前事实及过渡期记录同时存在记忆库中被检索，导致答案模型出错。研究提出A-TMA，一种状态感知覆盖层，保留被取代记录和过渡记录，构建与查询状态对应的证据包，并为QA步骤提供当前、历史和过渡标签。在冲突密集的LTP基准上，将A-TMA添加到Graphiti后，冲突准确率提升0.240绝对百分点。研究建议对持久化助手的记忆库、检索和答案分别评估。

## 正文

"Ghost memory" is a real problem with agents.

You might have seen the issue where a long-running agent still confidently repeats a user fact that stopped being true weeks ago?

New research names the failure "ghost memory."

Old facts, current facts, and the transition between them all sit in the memory bank at once, get retrieved together, and mislead the answer model. A-TMA is a state-aware overlay that keeps superseded and transition records instead of deleting them, builds evidence packets scoped to the state the query is asking about, and hands current, historical, and transition labels to the QA step.

Most memory benchmarks report only final QA accuracy, which hides where the error happened. On the conflict-heavy LTP benchmark, adding A-TMA to Graphiti lifts conflict accuracy by 0.240 absolute.

If you build persistent assistants, it's best to evaluate the bank, the retrieval, and the answer separately.

Paper: https://arxiv.org/abs/2607.01935

Learn to build effective AI agents in our academy: https://academy.dair.ai/