Source-linked AI summary
QUMem: Personalized Memory for Query-Conditioned User-State Inference in LLM Agents
Heng Wang, Yifei Li, Lingling Zhang, Pengyu Li, Xinyu Che, Xinyu Zhang, Zesheng Yang
TL;DR
Personalized memory systems must handle distributed, changing, and context-dependent user information, but existing boundaries, representations, and single-query retrieval can fail to preserve the evidence needed for valid user-state inference. QUMem builds semantic episodes, decomposes them into typed memories, and uses sequential query-conditioned retrieval and inference; it achieves the best reported performance on PersonaMem and KnowU-Bench, although reliable personalized execution remains challenging.
Problem
Existing personalized memory systems can disrupt event context, bind functionally different information together, and retrieve fragments that fail to jointly capture temporal validity and contextual applicability.
Method
QUMem constructs semantically continuous episodes, decomposes them into factual, preference, and transferable insight memories, and uses three agents for information needs, typed-memory retrieval planning, and user-state inference.
Results
QUMem achieves the best performance among evaluated methods on PersonaMem and KnowU-Bench, including a 4.6 percentage-point overall success-rate improvement over the strongest KnowU-Bench baseline.
Takeaways & Limitations
The results support preserving event context, distinguishing evidence roles, and performing task-driven user-state inference for long-term personalization.
Takeaways & Limitations
Experiments are otherwise conducted on PersonaMem with GPT-4o-mini as the base model.
Abstract
from arXiv · showhide
Large language model (LLM) agents increasingly use external memory systems to support personalization by drawing on long and evolving interaction histories, in which user preferences may be distributed across time, change with context, and conflict with earlier evidence. However, existing systems face three limitations: fixed-turn, fixed-token, or session-based boundaries can mix unrelated dialogue or split an event from its causes, decisions, and outcomes; storing multiple pieces of user information from the same interaction as a single memory binds together items that serve different functions and should be independently retrievable; and treating the current task as a single top-$k$ retrieval query can return fragments that are individually relevant but fail to jointly capture preference evolution, temporal validity, and contextual applicability. We introduce \textsc{QUMem}, a structured memory framework for query-conditioned user-state inference. \textsc{QUMem} first segments interaction histories into variable-length episodes according to semantic continuity, then decomposes each episode into independently retrievable factual, preference, and transferable insight memories while preserving temporal positions and source evidence. At inference time, three sequential agents identify task-specific information needs, plan multi-query retrieval over the typed memory stores, and jointly infer a temporally and contextually valid user state for downstream response generation. \textsc{QUMem} achieves state-of-the-art performance on both PersonaMem and KnowU-Bench, demonstrating the effectiveness of query-conditioned user-state inference for long-term personalization.
1 Introduction
QUMem addresses shortcomings in personalized memory by preserving event context, separating evidence by function, and inferring a temporally and contextually valid user state for each query. Its structured episodes, typed memories, and sequential agents support retrieval and joint interpretation of distributed user evidence.
- Personalized assistants must integrate distributed preferences, distinguish current from obsolete or context-dependent tendencies, and transfer prior decision rationales to new situations.
- Fixed turn, token, or session boundaries can mix unrelated dialogue or split events from their causes, decisions, feedback, and outcomes.
- Storing facts, preferences, and constraints from one interaction as an indivisible memory prevents retrieving only the information relevant to a particular query.
- Independent top-k retrieval can return individually relevant fragments that collectively miss state transitions, temporal validity, or the user’s current state.
- QUMem constructs semantically continuous episodes, decomposes them into factual, preference, and transferable insight memories, and preserves temporal positions and source links.
- Three sequential agents identify information needs, plan retrieval over typed memories, and infer the current state jointly supported by retrieved evidence.
2 Related Work
Related work spans structured personal-memory systems and retrieval methods that plan, route, filter, or synthesize evidence. QUMem differs by applying these ideas to typed interaction-history memories for query-conditioned user-state inference.
- A-MEM links structured experience notes, Mem0 manages salient information through explicit memory operations, and Zep tracks fact validity and conflicts with a temporal knowledge graph.
- Memory-structure research includes topic-aware segmentation with compressed retrieval, extending work on memory granularity and organization.
- Retrieval-augmented generation research uses query rewriting, pre-retrieval planning, and question decomposition to transform complex requests into retrieval-oriented queries or subquestions.
- Other retrieval systems select knowledge sources, generate source-specific queries, recursively route subquestions, or synthesize distributed evidence through explicit notes and specialized agents.
- QUMem operates over interaction-history memories rather than primarily addressing access to external knowledge.
3 Method
QUMem constructs adaptive dialogue episodes, decomposes them into typed atomic memories, and performs query-conditioned user-state inference through staged retrieval and evidence organization.
- 3.1 Problem Formulation: QUMem models a user’s chronologically ordered multi-session history and current query to produce a history-grounded user state for personalized output.The framework instantiates the memory function, while a downstream response model uses the inferred state with the query.
- 3.2 Dynamic Episode Construction: Semantic continuity determines variable-length episode boundaries, preserving the event context that fixed turn, token, or session boundaries can fragment or mix.A continuity classifier compares adjacent user utterances; assistant responses remain in the episode context but are excluded from the boundary decision.
- 3.3 Typed Memory Decomposition: Each episode is decomposed into independently retrievable factual, preference, and transferable insight memories, while retaining source episodes, supporting turns, and temporal positions.The three memory types respectively represent concrete experiences, context-specific choices or constraints, and evidence-grounded decision principles transferable to new contexts.
- 3.3 Typed Memory Decomposition: The typed memory system distinguishes factual experiences, preferences and constraints, and transferable decision principles as different functional roles of historical evidence.Preferences may reflect persistent tendencies or a particular stage in preference evolution, whereas insights abstract principles from concrete choices, feedback, and rationales.
- 3.4 Query-Conditioned User-State Inference: Three sequential agents identify information needs, plan rewritten queries over selected memory types, and infer the task-relevant user state from merged candidate memories.The retrieval plan selects one or more factual, preference, and insight stores for each rewritten query before candidates are merged and deduplicated.
4 Experiments
QUMem is evaluated on PersonaMem and KnowU-Bench against representative long-term memory baselines, with additional ablations, retrieval-depth sensitivity, and construction-cost analysis. It consistently improves benchmark performance, while its gains are strongest for distributed evidence and its efficiency comes from reducing construction triggers and bounding calls per construction.
- PersonaMem: QUMem outperforms all evaluated baselines across PersonaMem context configurations and query categories under both GPT-4o-mini and Gemini-3.5-flash.Overall accuracy rises from 52.99% to 61.02% with GPT-4o-mini and from 63.29% to 70.58% over the strongest baseline with Gemini-3.5-flash.
- PersonaMem: The largest PersonaMem gains occur when tracking preference evolution, making preference-aligned recommendations, and generalizing user information to new scenarios.These categories require jointly interpreting evidence from different times or contexts; gains are smaller for recognizing only the latest preference.
- PersonaMem: QUMem’s margin over the strongest baseline widens with longer histories, although absolute performance declines at 1M tokens and suggesting new ideas remains weakest.The results distinguish accurate preference inference from generating responses that are both novel and preference-aligned.
- KnowU-Bench: QUMem achieves the best KnowU-Bench performance across overall metrics and difficulty subsets, improving overall success rate by 4.6 percentage points over the strongest baseline.The benchmark tests translating historical preferences and constraints into concrete actions, but low success rates on hard tasks show that end-to-end personalized execution remains challenging.
- Ablation Studies: Removing any QUMem component degrades performance, with query-time User-State Reconstruction having the strongest overall effect and all component effects increasing for longer histories.Typed Memory Decomposition and Dynamic Episode Construction contribute by distinguishing evidence roles and preserving coherent event context.
- Ablation Studies: Retrieval depth k = 5 is the main setting because accuracy increases from 59.10% at k = 3 to 61.02% at k = 5, then slightly decreases to 60.78% at k = 10.The k = 5 setting achieves the highest overall accuracy and the best results in five of seven query categories.
5 Conclusion
QUMem combines dynamic episode construction, typed memory decomposition, and query-conditioned user-state inference to infer task-relevant user states from long, evolving histories. On PersonaMem, it achieves the best performance among evaluated methods.
- QUMem combines Dynamic Episode Construction, Typed Memory Decomposition, and Query-Conditioned User-State Inference for long-term personalization.
- QUMem achieves the best performance among the evaluated methods on PersonaMem.
- The framework preserves event context, distinguishes historical evidence roles, and performs task-driven user-state inference.