Paper 2511.03506
HaluMem: Evaluating Hallucinations in Memory Systems of Agents
- Published
- Nov 2025
- Research lab
- Independent
- Citations
- 37
- GitHub
- 151 stars
01 In brief
Summary
HaluMem is the first operation-level benchmark for evaluating hallucinations in memory systems for LLMs and AI agents.
It defines three tasks—memory extraction, memory updating, and memory question answering—to localize hallucinations to specific operational stages.
The benchmark includes two datasets, HaluMem-Medium and HaluMem-Long, with about 15k memory points and 3.5k questions each, featuring average dialogue lengths of 1.5k and 2.6k turns and context lengths up to 1M tokens.
A six-stage pipeline constructs user-centric, multi-turn human-AI dialogues with annotated memory points and questions.
Human annotation verified high quality (95.70% correctness).
Evaluations of systems like Mem0, Memobase, MemOS, Supermemory, and Zep revealed that most systems perform poorly on extraction and updating, with hallucinations accumulating and propagating to question answering.
MemOS performed best overall, while others declined significantly on the long-context dataset.
The study highlights the need for interpretable and constrained memory operations to suppress hallucinations and improve reliability.
02 From the paper
Abstract
Memory systems are key components that enable AI systems such as LLMs and AI agents to achieve long-term learning and sustained interaction. However, during memory storage and retrieval, these systems frequently exhibit memory hallucinations, including fabrication, errors, conflicts, and omissions. Existing evaluations of memory hallucinations are primarily end-to-end question answering, which makes it difficult to localize the operational stage within the memory system where hallucinations arise. To address this, we introduce the Hallucination in Memory Benchmark (HaluMem), the first operation level hallucination evaluation benchmark tailored to memory systems. HaluMem defines three evaluation tasks (memory extraction, memory updating, and memory question answering) to comprehensively reveal hallucination behaviors across different operational stages of interaction. To support evaluation, we construct user-centric, multi-turn human-AI interaction datasets, HaluMem-Medium and HaluMem-Long. Both include about 15k memory points and 3.5k multi-type questions. The average dialogue length per user reaches 1.5k and 2.6k turns, with context lengths exceeding 1M tokens, enabling evaluation of hallucinations across different context scales and task complexities. Empirical studies based on HaluMem show that existing memory systems tend to generate and accumulate hallucinations during the extraction and updating stages, which subsequently propagate errors to the question answering stage. Future research should focus on developing interpretable and constrained memory operation mechanisms that systematically suppress hallucinations and improve memory reliability.