Back to AI Research

AI Research

A-TMA: Decoupling State-Aware Memory Failures in Lo... | AI Research

Key Takeaways

  • A-TMA: Decoupling State-Aware Memory Failures in Long-Term Agent Memory Long-term memory allows AI agents to act as persistent assistants, but this creates a...
  • Long term memory lets LLM agents act as persistent assistants, but user facts change.
  • A useful memory system must know what is true now, what used to be true, and what changed.
  • We study \emph{ghost memory}, a state coordination failure in which old, current, and transition facts coexist in the memory bank, remain mixed during retrieval, and mislead the answer model.
  • We argue that memory systems should be understood and optimized from three levels: bank maintenance, retrieval, and answer time resolution.
Paper AbstractExpand

Long term memory lets LLM agents act as persistent assistants, but user facts change. A useful memory system must know what is true now, what used to be true, and what changed. We study \emph{ghost memory}, a state coordination failure in which old, current, and transition facts coexist in the memory bank, remain mixed during retrieval, and mislead the answer model. We argue that memory systems should be understood and optimized from three levels: bank maintenance, retrieval, and answer time resolution. We propose ATMA, a state aware overlay for existing memory systems. ATMA keeps superseded and transition records in the bank, builds evidence packets for the query's requested state view, and exposes current, historical, and transition labels to QA. We further call for decoupled evaluation of bank, retrieval, and answer level failures, since final QA accuracy can hide where ghost memory occurs. To make this failure measurable, we build LTP (LoCoMo Temporal Plus), a conflict heavy benchmark for ghost memory, and evaluate on LoCoMo for long conversation generalization. On LTP, Graphiti+ATMA improves conflict accuracy by 0.240 absolute over Graphiti. On LoCoMo, Graphiti+ATMA raises temporal F1 from 0.0295 to 0.1705. The gains are host dependent, but they indicate that explicit state roles can reduce memory failures hidden by final QA accuracy.

A-TMA: Decoupling State-Aware Memory Failures in Long-Term Agent Memory

Long-term memory allows AI agents to act as persistent assistants, but this creates a significant challenge: information about a user often changes over time. When an agent’s memory bank contains outdated facts alongside current ones, it can lead to "ghost memory"—a state coordination failure where the agent becomes confused by conflicting information. This paper introduces A-TMA, a state-aware overlay designed to help memory systems distinguish between current, historical, and transition facts to ensure agents provide accurate, context-aware responses.

The Problem of Ghost Memory

Ghost memory occurs when an agent’s memory bank holds multiple versions of the same fact—such as an old address and a new address—without clear labels. During retrieval, these mixed facts can mislead the AI, causing it to provide outdated or contradictory answers. The authors argue that to fix this, we must stop looking at memory as a single block and instead optimize it across three distinct levels: how the memory bank is maintained, how information is retrieved, and how the final answer is resolved.

How A-TMA Works

A-TMA acts as an intelligent layer that sits on top of existing memory systems. Instead of deleting old information, it keeps superseded and transition records in the bank but organizes them with specific labels. When a query is made, A-TMA builds "evidence packets" that provide the agent with a clear view of the requested state. By explicitly exposing labels like "current," "historical," or "transition" to the model, the system helps the AI understand the timeline of the information it is processing, rather than just seeing a jumble of conflicting data.

Measuring Success

To test this approach, the authors created a new benchmark called LTP (LoCoMo Temporal Plus), which is specifically designed to be "conflict-heavy" to expose ghost memory failures. They also evaluated the system on the existing LoCoMo benchmark to test long-term conversation generalization. The results showed significant improvements: when paired with the Graphiti memory system, A-TMA improved conflict accuracy by 0.240 on the LTP benchmark and increased the temporal F1 score on LoCoMo from 0.0295 to 0.1705.

Decoupling Evaluation

A key takeaway from this research is the call for "decoupled evaluation." The authors note that final question-answering accuracy can often hide the underlying causes of memory failures. By breaking down performance into bank, retrieval, and answer-level failures, researchers can better identify exactly where the memory system is breaking down. While the authors note that these gains can be host-dependent, the results demonstrate that explicitly labeling the roles of different memory states is a highly effective way to reduce the errors that typically plague long-term agent memory.

Comments (0)

No comments yet

Be the first to share your thoughts!