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

Sep 21, 2026

Uploaded by Aurora

Long term memory enables LLM agents to act as persistent assistants, but user facts can change, creating uncertainty about which state a query refers to: what is true now, what was true before, and what changed. The paper analyzes ghost memory, a state coordination failure where old, current, and transition facts coexist and remain mixed during retrieval, misleading QA. It proposes A-TMA, plus the LTP benchmark and LoCoMo evaluation, improving conflict accuracy and temporal F1.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages21

Reading Time53 min

AccessFree to download

Abstract

1 Introduction

Ghost memory and state coordination failure

Decoupling memory evaluation across bank, retrieval, and QA

Related approaches and limitations

Experimental results on LTP and LoCoMo

arXiv :2607 .0 1935v2 [ cs .AI] 8 Jul 2026 A-TMA: Decoupling State-Aware Memory Failures in Long-Term Agent Memory Zitong Shi1 Yixuan Tang1 ∗ Anthony Kum Hoe Tung1 1National University of Singapore [zitongshi@u.nus.edu](zitongshi@u.nus.edu) , [yixuan@comp.nus.edu.sg](yixuan@comp.nus.edu.sg) , [atung@comp.nus.edu.sg](atung@comp.nus.edu.sg) Abstract Long term memory lets LLM agents act as persistent assistants, but user facts change. The central question is not only whether a fact can be remembered, but whether the system can answer from the state the user is asking about: what is true now, what used to be true, and what changed. We study 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. The challenge is that simple deletion loses history, while timestamps or relevance scores alone do not expose which fact should be live for a given query. We therefore argue for decoupled memory evaluation at three levels: bank maintenance, retrieval, and answer time resolution. We propose A-TMA, a state aware overlay for existing memory systems that keeps superseded and transition records, builds evidence packets for the requested state view, and exposes current, historical, and transition labels to QA. 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/Zep +A-TMA improves conflict accuracy by 0.240 absolute over Graphiti/Zep. On LoCoMo, Graphiti/Zep +A-TMA 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. 1 Introduction LLM agents are moving from single turn chat toward persistent assistance. They remember user preferences, project decisions, constraints, and earlier conversations, then reuse this context when later requests depend on it. Agent systems already use memory for simulation, reflection, tool use, and open ended interaction [1–5] . Long term memory is therefore becoming infrastructure for agent systems [6–12] . Yet user memory is not a static profile. A user may change an address, leave a job, revise a plan, or ask about a past state. A memory system that treats every stored fact as timeless can be wrong even when each fact was true when it was written. We call this failure ghost memory. Ghost memory is not the mere presence of an old record. It is a state coordination failure. Old, current, and transition facts coexist in the bank, remain mixed during evidence construction, and reach the answer model without a clear state role. A former address should answer a question about where the user lived before. The same address should not answer where a package should be shipped now. A transition note can explain what changed. It should not be treated as the live value. The core challenge is therefore query conditioned state alignment: the system must preserve history, decide which state the query asks for, and expose that state to retrieval and QA. We study the following problem: how can state changing memory be evaluated by decoupling bank, ∗ Corresponding author. Preprint. Figure 1: Retrieval and QA state alignment. A historical query can retrieve both old and current values, but an uncontrolled packet may foreground the wrong state. Even with the right evidence present, unlabeled context can make QA follow the current value. A-TMA uses bounded retrieval control and explicit state labels to answer from the requested state. retrieval, and QA failures while preserving state history, retrieving evidence for the requested state view, and resolving answers with explicit state roles? Most long term memory systems address only part of this requirement [6–8, 13–18] . Profile and note based systems extract user facts and retrieve them

FAQs

What problem does the paper call ghost memory?

Ghost memory is a state coordination failure where old, current, and transition records coexist and stay mixed during evidence construction, reaching the answer model without a clear state role.

Why can simple deletion or timestamping be insufficient?

Simple deletion can lose history, while timestamps or relevance scores alone do not indicate which fact should be live for a specific query’s requested state.

How does A-TMA address state changes in long-term agent memory?

A-TMA decouples memory evaluation across bank maintenance, retrieval, and answer time resolution, keeps superseded and transition records, and provides explicit current, historical, and transition labels to QA.

Sep 21, 2026

Uploaded by Aurora