AuthorsChangjian Wang, Rongzhen Li, Weili Guan et al.
TL;DR
CueMem uses cue guided context reconstruction over a turn graph to reach 84.07% accuracy on LoCoMo, +5.26 points over MemoryOS.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
CueMem targets long term conversations where compressed memory units lose fine grained evidence and full histories cause latency and lost in the middle issues.
When agents answer questions like Which place we mentioned earlier would my mother prefer, coarse memory units mix signals and hide the exact supporting turn, breaking faithful question answering.
HOW IT WORKS
CueMem centers on Memory Cue Extraction, Turn Graph Construction, Context Reconstruction, and Memory Management to turn dialogue histories into cue guided, reconstructable memories.
You can think of CueMem like a human autobiographical memory: short cues act as prompts that lead you back through a graph of experiences instead of storing full transcripts.
This cue guided graph expansion lets CueMem recover precise dialogue evidence that a plain context window or monolithic memory units cannot expose without huge token and latency costs.
DIAGRAM
This diagram shows how CueMem processes a user query through cue retrieval, anchor selection, graph expansion, and answer generation at query time.
DIAGRAM
This diagram shows how CueMem is evaluated on LoCoMo and LongMemEval, including full history comparison and ablation variants.
PROCESS
Memory Cue Extraction
CueMem runs Memory Cue Extraction over each dialogue turn, using an LLM to produce relational triples and link them to source turns for later retrieval.
Turn Graph Construction
CueMem builds Turn Graph Construction by adding temporal edges within a window and semantic edges between turns whose cues are top k nearest neighbors in embedding space.
Context Reconstruction
At query time, CueMem performs Context Reconstruction by retrieving top m cues, treating their source turns as anchors, and expanding one hop over temporal and semantic edges.
Memory Management
CueMem applies Memory Management with simple ADD, DELETE, and UPDATE via new cues, plus a priority instruction to favor recent evidence during answer generation.
KEY CONTRIBUTIONS
Cue centered view of long term conversational memory
CueMem formalizes Memory Cue Extraction as fine grained relational triples linked to source turns, treating them as retrieval cues instead of self contained evidence for generation.
CueMem framework with graph based reconstruction
CueMem combines Turn Graph Construction and Context Reconstruction to expand from cue anchors over temporal and semantic edges, recovering compact yet faithful dialogue evidence.
Empirical gains on LoCoMo and LongMemEval
CueMem reaches 84.07% accuracy on LoCoMo and 75.20% on LongMemEval, surpassing strong baselines like MemoryOS and LightMem while using about 2K input tokens per query.
RESULTS
LoCoMo Overall
84.07%
+5.26 over MemoryOS
LoCoMo Single hop
75.18%
+1.42 over MemoryOS
LoCoMo Multi hop
82.24%
+4.05 over MemoryOS
LongMemEval Overall
75.20%
+1.60 over LightMem
CueMem is evaluated on LoCoMo and LongMemEval, which test single hop, multi hop, temporal, preference, and knowledge update questions. The gains show CueMem’s cue guided reconstruction improves accuracy over MemoryOS and LightMem while keeping contexts compact.
BENCHMARK
CueMem is evaluated on LoCoMo and LongMemEval, which test single hop, multi hop, temporal, preference, and knowledge update questions. The gains show CueMem’s cue guided reconstruction improves accuracy over MemoryOS and LightMem while keeping contexts compact.
BENCHMARK
Overall accuracy on LoCoMo for CueMem and representative long term memory baselines.
BENCHMARK
Overall accuracy on LongMemEval for CueMem and strong long term memory baselines.
KEY INSIGHT
CueMem without any explicit UPDATE operation still scores 87.18% on LongMemEval knowledge update questions, only 6.4 points below its full configuration.
This is surprising because many assume complex memory rewriting is necessary, but CueMem shows accurate reconstruction plus a simple recency priority instruction lets the LLM resolve conflicting facts.
WHY IT MATTERS
CueMem unlocks scalable long term conversational memory where agents reconstruct just enough context from cues and a turn graph instead of streaming entire histories.
Builders can now deploy agents that handle 100K token histories with about 2K token prompts, preserving fine grained evidence while keeping latency and cost low.
Guilin Zhang, Wei Jiang et al.
· 2026
A-MAC scores candidate memories using Utility, Confidence, Novelty, Recency, and Type Prior combined by a learned linear admission policy with Algorithm 1 A-MAC Memory Admission. On the LoCoMo benchmark, A-MAC achieves F1 0.583 and 2644 ms latency, improving F1 by 0.042 and reducing latency by 1187 ms compared to A-mem.
Robbyant Team, Zelin Gao et al.
arXiv 2026 · 2026
LingBot-World combines a Data Engine, Fundamental World Model, Action-Conditioned World Model, and Post-Training causal adaptation to turn a 28B-parameter video generator into a real-time interactive world simulator. On the VBench benchmark, LingBot-World achieves a dynamic degree of 0.8857 versus 0.7612 for Yume-1.5, while also improving imaging quality to 0.6683.
Manoj Madushanka Perera, Adnan Mahmood et al.
· 2026
AgenticAI-DialogGen chains ChatPreprocessor, KnowledgeExtractor, TopicAnalyzer, KnowledgeGraphBuilder, PersonaGenerator, DuelingChat Agent, ConversationValidator, ConversationRefiner, QAGeneration, and PostProcessing to turn raw multi-session chats into topic-guided, persona-grounded conversations with explicit short- and long-term memories. On the TGC / KG memory QA benchmark, Mistral-7B fine-tuned within AgenticAI-DialogGen achieves 87.36 F1, compared to GPT-4’s 83.77 F1 in a zero-shot setting on the same task.
Answers use this explainer on Memory Papers.
Checking…