Quang-Linh Tran

Lifelog Question-Answering (QA) has become an important application of lifelogging, answering questions posed over personal lifelog data to help lifeloggers gain insight into their lives, from lifestyle analytics to past-event reflection and behavioural patterns. Building on our previous work on lifelog retrieval, we present MemoriEase 4.0, our system for the Lifelog Search Challenge 2026 (LSC’26), which targets the QA task using reasoning large language models (LLMs). The system adopts a Retrieval-Augmented Generation (RAG) approach to answer questions. For complex questions that require aggregating information, such as counting or comparison across time, reasoning LLMs analyse the retrieved evidence step by step before producing an answer. At the live competition, MemoriEase 4.0 finished third overall among fifteen teams and recorded the strongest ad-hoc retrieval performance. Our analysis further identifies answer latency and aggregation accuracy as the principal challenges that remain for competitive lifelog QA. Beyond the challenge, the system aims to evolve into a complete assistant for everyday lifelog utilisation.

Lifelog retrieval plays an important role in memory support for lifeloggers. It helps the lifeloggers to browse, search and navigate their life moments from the lifelog data. However, the volume and variety of lifelog data are enormous and range in multiple modalities so they impose a big challenge to retrieve accurate lifelog moments. The Lifelog Search Challenges (LSCs) are a benchmark challenge for evaluating lifelog retrieval systems in different tasks. In this paper, we introduce the MemoriEase 2.0 lifelog retrieval system that participates in LSC'24. This system not only inherits core functions from the precedent system but also incorporates new components such as conversational search, visual similarity search and retrieval-augmented generation for question-answering tasks. The new functions are expected to help expert and novice users solve all topics in three tasks of LSC'24. We evaluate MemoriEase 2.0 in KIS topics in LSC'23 and the system achieves promising results with Recall@1 is 40% at the first hint and it solves 8 over 10 topics.

Lifelogging can be referred to as the process of passively collecting data on an individual's daily life. Lifelog data provides a large amount of information which can be used to understand the lifelogger's lifestyle and preferences. This data can also support the lifeloggers in saving their memories and important moments. Question-answering (QA) is a common task in natural language processing (NLP) and can be extended to multi-modal such as the visual question-answering task. QA for lifelog data can be described as the task of answering questions about a lifelogger's past using lifelog data, which can significantly help lifeloggers understand their life by asking questions about their lifelog. QA for lifelogs can also provide useful insights into lifelogger's life for those exploring their lifelog. This paper presents the MemoriQA lifelog dataset designed to explore the question-answering task for lifelogs. This dataset provides 61-day lifelog images and other lifelog data such as internet activity, health metrics, music listening history and GPS. A comprehensive annotation process is performed to create the description as well as question-answer pairs. We propose some methods to address the QA in lifelog problem in this paper.