AuthorsYijie Zhong, Yunfan Gao, Haofen Wang
TL;DR
HingeMem uses boundary-guided hyperedge memory plus query-adaptive retrieval to reach 63.9 F1 on LOCOMO, +5.5 over Zep without category templates.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
HingeMem targets long-term dialogue memory where performance drops around 30% when query categories are unspecified and fixed Top-k retrieval is used.
In such settings, systems like Memorybank and Zep lose 30–40% performance, causing unstable answers and inefficient, noisy retrieval over ultra-long histories.
HOW IT WORKS
HingeMem centers on Boundary Guided Long-Term Memory, Dialogue Boundary Extraction, Memory Construction, Query Adaptive Retrieval, Hyperedge Rerank, and Adaptive Stop to segment events and plan retrieval.
You can think of HingeMem like a hippocampus plus cortex: the cortex marks event boundaries, and the hippocampus stores hyperedges as an indexable card catalog of experiences.
This design lets HingeMem decide both what to retrieve and how much to retrieve, instead of stuffing a plain context window with a fixed Top-k list of memories.
DIAGRAM
This diagram shows how HingeMem analyzes a query, plans retrieval, reranks hyperedges, and adaptively stops to select memory.
DIAGRAM
This diagram shows how HingeMem is evaluated on LOCOMO and how the ablations compare boundary memory and adaptive retrieval.
PROCESS
Dialogue Boundary Extraction
HingeMem uses Dialogue Boundary Extraction to segment each session whenever person, time, location, or topic changes, producing element nodes and segment reasons.
Memory Construction
HingeMem runs Memory Construction to merge nodes, compute salience scores, cluster topics, and build hyperedges into Boundary Guided Long-Term Memory.
Query Adaptive Retrieval
Given a question, HingeMem applies Query Adaptive Retrieval to infer query type, select relevant elements, and generate a retrieval plan with element priorities.
Hyperedge Rerank and Adaptive Stop
HingeMem feeds candidate hyperedges into Hyperedge Rerank and Adaptive Stop, then passes the selected hyperedges as context for answer generation.
KEY CONTRIBUTIONS
Boundary Guided Long-Term Memory
HingeMem introduces Boundary Guided Long-Term Memory that writes hyperedges when person, time, location, or topic changes, preserving details while avoiding continuous summarization.
Query Adaptive Retrieval Mechanism
HingeMem proposes Query Adaptive Retrieval that predicts Recall Priority, Precision Priority, or Judgment queries and plans element-aware routing over the boundary memory.
Efficient Long-Term Memory on LOCOMO
HingeMem achieves 63.9 overall F1 and 75.1 LLM-as-a-Judge on LOCOMO, with 68% lower question answering token cost than HippoRAG2 while not using category-specific templates.
RESULTS
Overall F1
63.9
+7.0 over Zep (56.9 F1 without category format)
Overall J
75.1
+5.5 over HippoRAG2 (69.6 J with category format approximated from 70.6)
BLEU-1
0.404
+0.012 over Zep (0.392 BLEU-1 with category format)
Multi-Hop F1
53.6
+12.5 over HippoRAG2 (41.1 F1 with category format)
On the ultra-long dialogue benchmark LOCOMO, which averages 15,965.8 tokens per conversation and 1,986 questions across five categories, HingeMem demonstrates that boundary-guided memory plus query-adaptive retrieval can handle diverse query types without category templates. These results show that HingeMem scales across LLM sizes while improving both accuracy and efficiency for long-term conversational memory.
BENCHMARK
On the ultra-long dialogue benchmark LOCOMO, which averages 15,965.8 tokens per conversation and 1,986 questions across five categories, HingeMem demonstrates that boundary-guided memory plus query-adaptive retrieval can handle diverse query types without category templates. These results show that HingeMem scales across LLM sizes while improving both accuracy and efficiency for long-term conversational memory.
BENCHMARK
Overall F1 scores on LOCOMO when query categories are not provided to the systems.
BENCHMARK
Overall F1 for RAG with text memory versus HingeMem boundary memory and adaptive retrieval variants.
KEY INSIGHT
HingeMem without any category-specific question templates reaches 63.9 overall F1, while Zep with templates only gets 56.9 F1.
This is surprising because template-aware baselines should have an advantage, yet HingeMem’s boundary-guided memory and adaptive retrieval outperform them without extra category hints.
WHY IT MATTERS
HingeMem shows that boundary-triggered hyperedges plus query-adaptive retrieval can support scalable, interpretable long-term memory across ultra-long multi-session dialogues.
Builders can now deploy assistants that remember months of interaction, adapt retrieval depth per query type, and keep token costs manageable for web and edge applications.
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.
Yi Yu, Liuyi Yao et al.
arXiv 2026 · 2026
Agentic Memory (AgeMem) exposes memory management tools, a three-stage progressive RL strategy, and step-wise GRPO directly inside the agent policy to jointly control long-term and short-term memory. On Qwen3-4B-Instruct, AgeMem attains 54.31% average performance across ALFWorld, SciWorld, PDDL, BabyAI, and HotpotQA, exceeding the best baseline A-Mem at 45.74%.
Can Lv, Heng Chang et al.
· 2026
All-Mem organizes long-term agent memory through Online/Offline Decoupling, Agentic Topology Consolidation, and Topology-Aware Retrieval over a topology-structured memory bank. On LoCoMo, All-Mem reaches 54.63 4o-J versus Mem0’s 48.91, and on LongMemEval-S All-Mem reaches 60.20 4o-J versus Mem0’s 55.80.
Answers use this explainer on Memory Papers.
Checking…