PaperlayerView exact revision on arXiv

Source-linked AI summary

HaluMem: Evaluating Hallucinations in Memory Systems of Agents

Ding Chen, Simin Niu, Kehang Li, Peng Liu, Xiangping Zheng, Bo Tang, Xinchi Li, Feiyu Xiong, Zhiyu Li

arXiv:2511.03506v3cs.CL

TL;DR

Memory-system hallucinations are difficult to localize because prior evaluations mainly use end-to-end question answering. HaluMem addresses this gap with operation-level tasks and user-centered long-context benchmarks, finding persistent bottlenecks and cumulative hallucination effects across memory operations.

  • Problem

    Existing end-to-end question-answering evaluations make it difficult to identify whether memory hallucinations arise during extraction, updating, or question answering.

  • Method

    HaluMem evaluates memory extraction, memory updating, and memory question answering using user-centered multi-turn HaluMem-Medium and HaluMem-Long datasets.

  • Results

    Stage-wise evaluation reveals persistent bottlenecks and cumulative, amplifying hallucination effects across memory extraction, updating, and question answering.

  • Takeaways & Limitations

    The findings motivate interpretable and constrained memory-operation mechanisms that improve extraction, updating, robustness, and question-answering reliability.

  • Takeaways & Limitations

    Zep's lack of a Get Dialogue Memory API prevents computation of memory-extraction metrics.

Abstract

from arXiv · show

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.

1 Introduction

Memory systems support personalized, coherent long-term interaction by organizing and updating user information, but hallucinations can arise across extraction, updating, and question answering. HaluMem introduces operation-level evaluation with user-centered multi-turn benchmarks to localize these behaviors and examine their accumulation.

  • Memory systems record, structure, update, and retrieve user information to support personalization and consistency across long-term human–AI conversations.
  • Memory hallucinations include fabricated, erroneous, conflicting, or missing information, and upstream errors can be amplified during generation.
  • HaluMem is the first operation-level benchmark evaluating hallucinations across memory extraction, memory updating, and memory-based question answering.
  • The benchmark enables fine-grained diagnosis by measuring extraction accuracy and coverage, updating consistency, and question-answering errors from incorrect references or fabricated content.
  • HaluMem-Medium and HaluMem-Long each contain approximately 15,000 memory points and more than 3,400 queries, with HaluMem-Long extending contexts to millions of tokens.
  • Stage-wise evaluation reveals cumulative and amplifying hallucination effects across extraction, updating, and question answering.

2 Related Work

Prior memory systems emphasize manageable external storage and retrieval, but hallucination evaluation remains largely end-to-end and difficult to localize. HaluMem addresses this gap with operation-level analysis of memory hallucinations.

  • 2.1 Memory System: Existing memory systems include external-memory approaches such as RAG, GraphRAG, and user-level long-term memory systems.RAG improves manageability through editable plaintext retrieval, while GraphRAG adds entity–relation structure for relational reasoning.
  • 2.2 Evaluation Hallucinations in Memory Systems: Memory hallucinations include fabricated, erroneous, conflicting, outdated, or missing information during storage, updating, and retrieval.They can act as upstream causes of generation hallucinations and may be amplified or obscured by generated responses.
  • 2.2 Evaluation Hallucinations in Memory Systems: Existing benchmarks such as LoCoMo, LongMemEval, PrefEval, and PersonaMem primarily assess overall memory performance through end-to-end tasks.These benchmarks evaluate outcomes such as recall, event tracking, summarization, or generation rather than directly measuring hallucination sources.
  • 2.2 Evaluation Hallucinations in Memory Systems: HaluMem is designed for fine-grained, operation-level hallucination analysis, filling a gap left by black-box memory evaluations.Its focus is to attribute errors to specific memory operations rather than infer them only from final task performance.

3 Problem Definition

The problem definition separates memory-system processing into extraction, updating, retrieval, and question answering, while identifying the limits of end-to-end answer metrics. Stage-specific gold standards make hallucination sources more traceable.

  • 3 Problem Definition: HaluMem models memory processing with extraction E, updating U, retrieval R, and question answering Q.Extraction adds memory points, updating modifies or deletes existing memories, retrieval selects relevant memories, and question answering generates responses.
  • 3 Problem Definition: End-to-end answer accuracy cannot identify whether an error originated in extraction, updating, or question answering.This lack of traceability limits the development of targeted mitigation strategies.
  • 3 Problem Definition: The benchmark evaluates extraction, updating, and question answering with stage-specific gold standards and metrics.The gold standards include newly added memory points, before–after update pairs, and gold answers for queries.

4 Methodology for Constructing HaluMem

HaluMem is constructed through a user-centered, six-stage progressive-expansion pipeline that turns evolving personas into adversarial multi-turn dialogues and memory questions. Human annotation verifies dataset quality, producing Medium and Long variants at different context scales.

  • 4 Methodology for Constructing HaluMem: The construction pipeline uses six stages to generate user-centered, controllable datasets through progressive expansion.The stages cover persona construction, life-skeleton design, event-flow creation, session and memory-point generation, dialogue generation, and question generation.
  • 4 Methodology for Constructing HaluMem: The event-flow stage converts persona evolution into chronological memory timelines combining initial, career, and daily preference events.Career events provide the main storyline, while preference changes and initial information add contextual detail.
  • 4 Methodology for Constructing HaluMem: Session generation creates context-rich, goal-driven, adversarial dialogues and applies memory self-verification for consistency.Adversarial injection adds distractor memories before multi-turn dialogue generation and verification.
  • 4 Methodology for Constructing HaluMem: Question generation creates balanced memory-related question–answer pairs across six predefined categories and varying event complexity.Career-event substages are integrated into single units to increase reasoning depth.
  • 4 Methodology for Constructing HaluMem: 95.70% correctness was reported in human annotation of 700 HaluMem-Medium sessions covering memory points and question–answer pairs.Eight annotators evaluated correctness, relevance, and consistency across randomly selected sessions.
  • 4 Methodology for Constructing HaluMem: HaluMem-Medium contains 30,073 dialogue rounds, 14,948 memory points, and 3,467 QA pairs, while HaluMem-Long extends contexts to 1M tokens.The Medium dataset covers 20 users with about 160k-token average contexts; Long contains 53,516 dialogue rounds.

5 Evaluation Framework of HaluMem

HaluMem evaluates memory systems session by session across extraction, updating, and question answering, aggregating task metrics after all chronological interactions. Its framework combines APIs, annotated memory operations, and metrics for completeness, correctness, omissions, hallucinations, and answer quality.

  • HaluMem feeds each user’s dialogue sessions chronologically into the memory system, triggers the relevant evaluation after each session, and aggregates the three task metrics.
  • The framework provides Add Dialogue, Get Dialogue Memory, and Retrieve Memory APIs to support extraction, inspection, and query-based memory retrieval.
  • 5.1 Memory Extraction: Memory extraction measures whether systems store key dialogue information without fabrication or irrelevance, using completeness, correctness, and false-memory resistance metrics.
  • 5.1 Memory Extraction: The extraction F1 score jointly evaluates memory recall and target-memory precision, combining completeness and correctness into one measure.
  • 5.2 Memory Updating: Memory updating evaluates correct modification, merging, or replacement of existing memories while tracking incorrect updates, omissions, and conflicts.
  • 5.3 Memory Question Answering: Memory question answering evaluates the end-to-end pipeline by retrieving relevant memories, generating answers, and comparing them with references using correctness, hallucination, and omission measures.

6 Experiments

Experiments evaluate multiple memory systems across HaluMem’s extraction, updating, and question-answering tasks, with additional analyses of memory types, question categories, and runtime. Results show substantial degradation on long contexts, bottlenecks in updating, limited QA accuracy, and high computational costs during memory writing.

  • Overall Evaluation on HaluMem: Most systems perform worse on HaluMem-Long than on HaluMem-Medium, with Mem0, Mem0-Graph, and Memobase showing particularly notable declines.Supermemory and MemOS show the opposite extraction trend, while the declines indicate difficulty processing irrelevant information and distinguishing high- from low-value memories.
  • Overall Evaluation on HaluMem: Except for MemOS, all systems achieve memory-integrity recall rates below 60%, while all systems have accuracy below 62%.The extraction results indicate missing reference memories and a high proportion of hallucinations; MemOS and Supermemory perform best on F1 and remain stable in long contexts.
  • Overall Evaluation on HaluMem: Most systems suffer omission rates above 50% during memory updating, and performance drops considerably on HaluMem-Long.The results identify weak linkage between extraction and updating as a bottleneck, while low hallucination rates can reflect few samples entering the update stage.
  • Overall Evaluation on HaluMem: All systems achieve answer accuracies below 70%, with hallucination and omission rates remaining high and overall performance further decreasing on HaluMem-Long.The reported results connect QA performance with the sufficiency and accuracy of upstream memory extraction under interference and extended contexts.
  • Typewise Memory Evaluation: MemOS achieves the best overall typewise extraction performance, but all systems remain weak across event, persona, and relationship memories.Mem0, Mem0-Graph, and Memobase decline in long-context scenarios, while personal traits are easier to capture than event dynamics and relationship changes.
  • Performance on Different Question Types: Performance is relatively better on memory-boundary and conflict questions but deteriorates on multi-hop inference, dynamic update, and generalization questions.MemOS, SuperMemory, and Zep remain relatively stable across datasets, whereas the Mem0 series and Memobase degrade on HaluMem-Long.
  • Efficiency Analysis of Memory Systems: Dialogue addition requires substantially more time than memory retrieval, making the write stage the primary computational bottleneck.On HaluMem-Medium, Supermemory has the best dialogue-addition and total runtime, MemOS has the best retrieval efficiency, and Mem0 variants exceed 2700 minutes for dialogue addition.

7 Conclusion

HaluMem addresses the limitations of black-box, end-to-end memory evaluation by assessing hallucinations at three operational levels. Its experiments expose persistent weaknesses across memory extraction, updating, and question answering, motivating improvements in memory operation quality and efficiency.

  • HaluMem evaluates memory extraction, memory updating, and memory question answering to analyze hallucinations across operational levels.
  • The benchmark reveals persistent bottlenecks in coverage, accuracy, update capability, robustness to interference, and question answering reliability.
  • Future work should improve extraction quality, update logic, semantic understanding, and system efficiency for more stable and comprehensive long-term memory.

Appendices

The appendices define HaluMem’s memory and question taxonomies for systematic hallucination evaluation. They distinguish three memory types and six question types covering recall, reasoning, updates, boundaries, application, and conflict correction.

  • HaluMem organizes memory content into persona, event, and relationship memories, representing different semantic levels and stability characteristics.
  • The benchmark includes six question types: basic fact recall, multi-hop inference, dynamic update, memory boundary, generalization and application, and memory conflict.
  • Memory-boundary questions test whether systems identify unmentioned information instead of fabricating answers.
  • Memory-conflict questions test whether systems detect contradictions, correct erroneous premises, and answer using correct memory.

A.3 Dataset Statistics

HaluMem provides finalized statistical documentation for medium- and long-context datasets. HaluMem-Long extends HaluMem-Medium with irrelevant dialogue inserted within and between sessions to test robustness under ultra-long contexts.

  • The statistical overview covers context scale, session quantity, memory distribution, and question-type composition for both finalized datasets.
  • HaluMem-Long builds on HaluMem-Medium and targets robustness and hallucination suppression in ultra-long-context scenarios.
  • Within-session irrelevant exchanges and between-session irrelevant dialogue blocks are inserted into HaluMem-Long.
  • Inserted irrelevant dialogue includes factual question-answer pairs and mathematical-reasoning question-answer pairs.
  • Additional factual question-answer pairs were sampled across eight domains using GPT-4o to increase irrelevant-dialogue diversity.

B Special Configurations for Some Memory Systems

The appendix documents system-specific configuration adjustments made for reproducibility when memory-system APIs impose constraints. For Memobase, localized database access and task-specific recall limits compensate for unavailable or restricted APIs.

  • Some evaluated memory systems require configuration workarounds because of unique API constraints.
  • Memobase memories were accessed directly from its underlying database because the system lacks a Get Dialogue Memory API.
  • Memobase recall limits were set to 250 tokens for memory updating and 500 tokens for memory question answering.

B.2 Zep

Zep could not be evaluated for memory extraction because its official APIs did not expose all memories from a session. The section also describes the annotation and scoring framework used for memory and question-answering items.

  • B.2 Zep: Zep’s official APIs did not provide complete session-level memory retrieval, preventing evaluation of its memory extraction performance.The available thread method returned only recent memories rather than the complete set required by the evaluation.
  • B.2 Zep: Zep’s asynchronous workflow also prevented accurate measurement of dialogue-addition time, so only memory-retrieval cost was recorded.This limitation concerns timing measurement rather than the completeness of retrieved memories.
  • B.2 Zep: The annotation dataset combines user persona information, multi-turn dialogue, and evaluation items consisting of memory points or question-answer pairs.Items are labeled by evaluation type and categorized into memory or question types such as persona, event, relationship, and basic fact recall.
  • B.2 Zep: Question categories test direct recall, multi-hop inference, dynamic updates, memory boundaries, generalization, and conflicts requiring correction of erroneous premises.The categories span straightforward retrieval, reasoning across dialogue, temporal changes, unknown information, preference-based applications, and contradiction handling.
  • B.2 Zep: Memory points and question-answer pairs are judged for correctness, relevance, and consistency using binary decisions and 0–10 dimension scores.Consistency and relevance are each mapped to low, partial, or high ranges from 0–3, 4–6, and 7–10.

D Prompts

The appendix documents prompts for memory question answering and for scoring extraction, updating, and question-answering outputs. It also gives annotation criteria and examples of the structured user profiles used in evaluation.

  • D Prompts: The prompt appendix covers memory question answering and scoring prompts for memory extraction, updating, and question answering.The question-answering templates retrieve memories before GPT-4o generates responses, while separate templates guide scoring across the three tasks.
  • D.1 Prompts for Memory Question Answering Task: Question-answering prompts instruct systems to analyze timestamped memories, use direct evidence, and prioritize the most recent memory when entries conflict.Some templates also require synthesizing information across multiple memory entries.
  • D.2 Prompts for Scoring in Memory Evaluation Tasks: The scoring prompts evaluate memory integrity, memory accuracy, update accuracy, and response labels such as Correct, Hallucination, and Omission.Memory integrity focuses on missed key points, while accuracy scoring checks candidate memories against dialogue and golden memories.
  • D.2 Prompts for Scoring in Memory Evaluation Tasks: Memory accuracy scoring assigns 2 points to fully supported memories, 1 to partially supported memories, and 0 to unsupported or contradictory memories.The procedure decomposes each candidate into atomic information points and checks each against dialogue and golden memories.
  • D.2 Prompts for Scoring in Memory Evaluation Tasks: Question-answering scoring treats missing elements as Omission and definite answers to unknown references as Hallucination.Equivalent numerical expressions are allowed only when their underlying numerical values remain unchanged.
  • E.1 User Profile Example in Stage 1: The annotation examples distinguish core profile, dynamic state, and preference information in structured JSON generated during stage 1.The core profile example includes fields such as birth date, degree, major, family status, and relationship status, alongside parent records.
  • E.1 User Profile Example in Stage 1: The profile examples include parent relationship types, birth dates, and descriptions, illustrating how family information is represented in structured records.The shown records include father and mother entries with associated descriptions.

E.4 Examples of irrelevant dialogues

The appendix illustrates irrelevant dialogues with examples spanning general knowledge, chemistry, and mathematics rather than user-specific information.

  • E.4 Examples of irrelevant dialogues: The examples classify dialogues about Thomas Edison, bleach concentration, and right-triangle geometry as irrelevant dialogues.These exchanges concern general factual, scientific, or mathematical questions rather than user-centered memory content.
  • E.4 Examples of irrelevant dialogues: The bleach example asks for the pH of concentrated bleach and includes a concentration explanation in grams per 100 milliliters.The example uses 8.25% w/v as the stated concentration representation.
  • E.4 Examples of irrelevant dialogues: The mathematical example asks for a locus equation involving a right triangle and squared distances, illustrating a non-personal problem-solving exchange.The dialogue begins by specifying triangle coordinates before outlining the derivation.
Loading 2511.03506v3…