MemUse: Moving Memory Evaluation from Direct QA to Natural Integration in Long-Term Human-AI Conversation

Hugging Face Daily Papers Papers

Summary

The paper introduces MemUse, a benchmark that shows direct QA accuracy for conversational memory does not predict user satisfaction, whereas natural integration of prior context does, revealing a large gap between recall and conversational use.

Memory systems for conversational LLMs are conventionally evaluated by direct, fact-seeking questions about prior dialogue (Direct QA): can the model recall fact X from a prior conversation? We tested whether higher Direct QA accuracy correlates with higher user satisfaction in a 4-month deployment (40 users, 1,872 sessions, 7 memory conditions). Existing-benchmark Direct QA varies from 19.7% to 70.1% across the 7 conditions, but satisfaction does not change. We hypothesize that existing benchmarks and user satisfaction are tracking different capabilities: benchmarks measure elicited retrieval (recall when asked), while conversation requires natural integration (detecting relevance and naturally weaving prior context into a response). To examine this, we introduce MemUse, a set of real user-cued memory moments drawn from the deployment, scored by an integration-aware judgment of the natural conversational response. Holding the model and context fixed, the same system that scores 78.8% on Direct QA references only 7.9% of those facts in conversation -- a 71-point gap. Within these moments, Natural Integration is associated with satisfaction, whereas Direct QA is not. We release the deployment corpus and MemUse together with all judgments and scoring prompts at https://github.com/ryuichi-sumida/memuse.
Original Article
View Cached Full Text

Cached at: 08/27/26, 03:19 AM

Paper page - MemUse: Moving Memory Evaluation from Direct QA to Natural Integration in Long-Term Human-AI Conversation

Source: https://huggingface.co/papers/2608.24189

Abstract

Direct QA benchmarks for conversational memory do not predict user satisfaction, whereas natural integration of prior context does, revealing a large gap between elicited recall and conversational use.

Memory systemsforconversational LLMsare conventionally evaluated by direct, fact-seeking questions about prior dialogue (Direct QA): can the model recall fact X from a prior conversation? We tested whether higherDirect QAaccuracy correlates with higher user satisfaction in a 4-month deployment (40 users, 1,872 sessions, 7 memory conditions). Existing-benchmarkDirect QAvaries from 19.7% to 70.1% across the 7 conditions, but satisfaction does not change. We hypothesize that existing benchmarks and user satisfaction are tracking different capabilities: benchmarks measureelicited retrieval(recall when asked), while conversation requiresnatural integration(detecting relevance and naturally weaving prior context into a response). To examine this, we introduceMemUse, a set of real user-cued memory moments drawn from the deployment, scored by an integration-aware judgment of the natural conversational response. Holding the model and context fixed, the same system that scores 78.8% onDirect QAreferences only 7.9% of those facts in conversation -- a 71-point gap. Within these moments,Natural Integrationis associated with satisfaction, whereasDirect QAis not. We release the deployment corpus andMemUsetogether with all judgments and scoring prompts at https://github.com/ryuichi-sumida/memuse.

View arXiv pageView PDFGitHub0Add to collection

Community

Paper author

Paper submitter

about 2 hours ago

Accepted to EMNLP 2026 Main.

We ran a 4-month real-world deployment of a memory-augmented conversational AI and found that Direct QA accuracy - the standard way to evaluate long-term memory - doesn’t predict user satisfaction: a system scoring 78.8% on Direct QA spontaneously referenced only 7.9% of those facts in actual conversation.

We introduce MemUse, a benchmark built from real user interactions that instead measures whether models naturally integrate remembered context into responses.

Upload images, audio, and videos by dragging in the text input, pasting, orclicking here.

Tap or paste here to upload images

Get this paper in your agent:

hf papers read 2608\.24189

Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash

Models citing this paper0

No model linking this paper

Cite arxiv.org/abs/2608.24189 in a model README.md to link it from this page.

Datasets citing this paper1

#### RuiSumida/memuse Viewer• Updatedabout 2 hours ago • 2.06k • 64

Spaces citing this paper0

No Space linking this paper

Cite arxiv.org/abs/2608.24189 in a Space README.md to link it from this page.

Collections including this paper0

No Collection including this paper

Add this paper to acollectionto link it from this page.

Similar Articles

MemoryDocDataSet: A Benchmark for Joint Conversational Memory and Long Document Reasoning

arXiv cs.CL

MemoryDocDataSet is a new synthetic benchmark of 50 micro-worlds and 1,000 QA pairs designed to evaluate AI systems on the joint task of conversational memory and long-document reasoning simultaneously. The best baseline (RAG-Both) achieves only 0.358 overall F1, highlighting a significant gap in current systems' ability to unify conversational memory with long-document navigation.

SuperMemory-VQA: An Egocentric Visual Question-Answering Benchmark for Long-Horizon Memory

Hugging Face Daily Papers

SuperMemory-VQA is a new egocentric VQA benchmark featuring 52.9 hours of AI-glasses footage and 4,853 QA pairs designed to evaluate AI assistants on long-horizon memory tasks spanning object recall, intent, timelines, and conversations. Benchmarking reveals existing agentic frameworks and LLMs remain far from reliable on these real-world memory challenges.