Is Your AI Agent Still Repeating What You Said Last Week? This Is “Ghost Memory”
Long-running AI agents stubbornly repeat facts that users mentioned weeks ago, even after those facts have become outdated. A new study from the National University of Singapore has named this failure mode Ghost Memory, and proposed the A-TMA solution that enables memory systems to learn to distinguish between “present”, “past”, and “changes”.
If you've ever built a persistent AI assistant, you've almost certainly encountered this scenario: an agent that has been running for weeks still confidently repeats a fact the user told it three weeks ago, even though that fact has long since changed.
This is not an isolated case. Researchers Zitong Shi, Yixuan Tang, and Anthony Kum Hoe Tung from the National University of Singapore have named this failure mode in their latest paper — **Ghost Memory**.
Where does the problem lie? Outdated facts, current facts, and the transition states between the two are all mixed together in the memory bank. They are all retrieved together during querying and passed to the generative model. When the model sees two conflicting pieces of information, it picks the outdated one and outputs it confidently.
Their solution is called **A-TMA**, a state-aware overlay. The core idea is straightforward:
- Instead of deleting superseded records, retain them and mark them as “outdated”.
- Record the transition event “changed from A to B”.
- During retrieval, construct a evidence package that only includes information relevant to the temporal state of the query.
- When passing the evidence to the QA model, explicitly tag it as “current”, “historical”, or “transitional”.
What was the result? On the conflict-intensive LTP benchmark, adding A-TMA to Graphiti boosted the conflict accuracy by 0.240 in absolute terms.
One reply put it perfectly: “Ghost memory isn't a retrieval bug. Storing facts and learning the truth are two different things. A memory system that cannot represent 'this used to be true' isn't memory — it's a delusional cache.”
Another reply was even more straightforward: “Memory needs expiration semantics.”
Paper link: https://arxiv.org/abs/2607.01935
If you are also building a persistent assistant, it's recommended to split your evaluation into three layers: the memory bank itself, the retrieval logic, and the answer model. Only looking at the final QA accuracy will hide where ghost memory actually occurs.
After all, an AI that can only remember but never forget is no different from a human-shaped tape recorder.
发布时间: 2026-07-07 02:32