ATMA: A State-Aware Memory Overlay to Resolve Ghost Memory Failures in LLM Agents
By
[Submitted on 2 Jul 2026]
Summary
This paper introduces "ghost memory" — a state coordination failure in LLM agent memory systems where old, current, and transition facts coexist and get mixed during retrieval, misleading the answer model. The authors propose ATMA (A-TMA), a state-aware overlay for existing memory systems that keeps superseded and transition records, builds evidence packets for requested state views, and exposes current/historical/transition labels to QA. They also build LTP (LoCoMo Temporal Plus), a conflict-heavy benchmark for ghost memory evaluation. Results show Graphiti+ATMA improves conflict accuracy by 0.240 on LTP and raises temporal F1 from 0.0295 to 0.1705 on LoCoMo.
Source
Twitter / XATMA: A State-Aware Memory Overlay to Resolve Ghost Memory Failures in LLM Agentsarxiv.orgKey quotes
· 5 pulledWe 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.
We argue that memory systems should be understood and optimized from three levels: bank maintenance, retrieval, and answer time resolution.
We further call for decoupled evaluation of bank, retrieval, and answer level failures, since final QA accuracy can hide where ghost memory occurs.
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.
You might also wanna read
Memori launches agent-native persistent memory infrastructure using structured knowledge graphs from agent trace data
Memori is a new agent-native memory infrastructure that enables AI agents to create structured, long-term persistent memory directly from ag
Product Hunt·2mo ago
Agent Memory Is Distributed State Management, Not Magic
The article argues that "agent memory" in AI systems is fundamentally just distributed state management rebranded. It draws parallels betwee
Building Production AI Agents: A Multi-Tiered Memory Architecture with Zep, Mem0, and ContextNest
This article discusses the need for a multi-tiered persistent memory architecture when building production-grade AI agents. It argues that r
promptowl.ai·4d ago
Research Directions for Overcoming Memory and Interconnect Challenges in Large Language Model Inference Hardware
This article discusses the technical challenges of Large Language Model (LLM) inference, highlighting how the autoregressive Decode phase ma
arxiv.org·5mo ago
Multi-Stream LLMs: A Parallel Architecture to Overcome Single-Stream Bottlenecks in Language Models
This paper introduces "Multi-Stream LLMs," a novel approach to overcoming the limitations of current language model architectures that rely
arxiv.org·1mo ago
Comments
Sign in to join the conversation.
No comments yet. Be the first.