AuthorsHeng Wang, Yifei Li, Lingling Zhang et al.
TL;DR
QUMem uses Dynamic Episode Construction plus Typed Memory Decomposition and query-conditioned user-state inference to reach 70.58% overall accuracy on PersonaMem with Gemini-3.5-flash, +7.29 points over Mem0.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
QUMem targets memory units where fixed-turn or fixed-token boundaries mix unrelated dialogue or split events, making later retrieval unable to repair disrupted context.
These fragmented memories cause long-term assistant agents to misinterpret preferences, ignore temporal validity, and apply obsolete or context-specific constraints in current tasks.
HOW IT WORKS
QUMem’s core mechanism combines Dynamic Episode Construction, Typed Memory Decomposition, and Query-Conditioned User-State Inference to preserve event context and separate factual, preference, and transferable insight memories.
You can think of QUMem as a card catalog where each event becomes a folder, and each folder holds separate cards for facts, preferences, and reusable decision principles.
This design lets QUMem retrieve and jointly interpret distributed evidence, enabling temporally valid user-state inference that a single flat context window or top k similarity search cannot provide.
DIAGRAM
This diagram shows how QUMem’s three agents transform a user query into a structured, query-conditioned user state using typed memory stores.
DIAGRAM
This diagram shows how QUMem is evaluated on PersonaMem and KnowU-Bench, including component ablations and different context lengths.
PROCESS
Dynamic Episode Construction
QUMem uses Dynamic Episode Construction with a semantic continuity classifier to group adjacent user utterances into variable length dialogue episodes that preserve event context.
Typed Memory Decomposition
Within each episode, QUMem applies Typed Memory Decomposition to extract atomic factual, preference, and transferable insight memories with temporal positions and provenance.
Query-Conditioned User-State Inference
QUMem’s Information Need Agent, Retrieval Planning Agent, and User-State Inference Agent collaborate to identify information needs, plan multi query retrieval, and reconstruct Zq.
Personalized Response Generation
Given the structured user state Zq, QUMem feeds Fq, Tq, and Iq into the downstream LLM agent Ψ to generate temporally valid, preference aligned personalized responses.
KEY CONTRIBUTIONS
Structured Memory Framework for Long Term Personalization
QUMem introduces Dynamic Episode Construction and Typed Memory Decomposition to build factual, preference, and transferable insight memories while preserving event context and temporal positions.
Query Conditioned Three Agent User State Inference
QUMem’s Information Need Agent, Retrieval Planning Agent, and User State Inference Agent separate evidence requirements, retrieval planning, and state reconstruction for query conditioned user modeling.
State of the Art on PersonaMem and KnowU Bench
QUMem reaches 61.02% overall accuracy with GPT 4o mini and 70.58% with Gemini 3.5 flash on PersonaMem, and achieves 17.4% success rate on KnowU Bench, surpassing Mem0 and Zep.
RESULTS
Overall
61.02%
+8.03 over Mem0 with GPT-4o-mini
Track full preference evolution
61.13%
+10.49 over Mem0 with GPT-4o-mini
Provide preference-aligned recommendation
66.96%
+11.26 over Mem0 with GPT-4o-mini
Success Rate
17.4%
+4.6 over Mem0 on KnowU-Bench overall
On PersonaMem, which tests dynamic user modeling and preference evolution, QUMem consistently leads A-MEM, Mem0, and Zep across context lengths and query categories. On KnowU-Bench, which evaluates personalized mobile agents, QUMem’s 17.4% success rate shows that query-conditioned user-state inference improves concrete task execution.
BENCHMARK
On PersonaMem, which tests dynamic user modeling and preference evolution, QUMem consistently leads A-MEM, Mem0, and Zep across context lengths and query categories. On KnowU-Bench, which evaluates personalized mobile agents, QUMem’s 17.4% success rate shows that query-conditioned user-state inference improves concrete task execution.
BENCHMARK
Overall accuracy (%) on PersonaMem with Gemini-3.5-flash across all context lengths.
BENCHMARK
Overall accuracy (%) on PersonaMem with GPT-4o-mini across all context lengths for QUMem ablations.
KEY INSIGHT
QUMem’s overall accuracy on PersonaMem actually increases from 59.10% at k equals 3 to 61.02% at k equals 5, but then drops to 60.78% at k equals 10.
This is surprising because many practitioners assume more retrieved memories always help, yet QUMem shows that excessive candidates can dilute relevance and slightly hurt user-state inference.
WHY IT MATTERS
QUMem unlocks query-conditioned user-state inference that tracks full preference evolution, revisits update reasons, and generalizes decision principles to new scenarios.
Builders can now design long-lived LLM agents that maintain temporally valid, provenance-aware user profiles instead of relying on brittle top k retrieval over flat, session-based memory chunks.
Nikhil Verma
· 2026
Focus Agent adds start_focus, complete_focus, a persistent Knowledge block, and an optimized Persistent Bash plus String-Replace Editor scaffold to actively compress context during long software-engineering tasks. On five hard SWE-bench Lite instances against a Baseline ReAct agent, Focus Agent achieves 22.7% token reduction (14.9M → 11.5M) while matching 3/5 = 60% task success.
Xingyu Lyu, Jianfeng He et al.
· 2026
ADAM combines Anchor extraction, Distribution estimation, Anchor selection, and Query generation to adaptively probe agent memory via an auxiliary generator and entropy based selection. On the EHRAgent benchmark with Llama2-7b-chat, ADAM reaches EQ=77 and ASR=1.00, compared to MEXTRA’s EQ=44 and ASR=0.89.
Shannan Yan, Jingchen Ni et al.
· 2026
AdaMem organizes dialogue history into Working Memory, Episodic Memory, Persona Memory, and Graph Memory coordinated by a Memory Agent, Research Agent, and Working Agent. On LoCoMo with GPT-4.1-mini, AdaMem achieves 44.65 F1 overall, beating the best baseline LangMem at 41.76 F1 by +2.89.
Answers use this explainer on Memory Papers.
Checking…