AuthorsNingning Zhang, Xingxing Yang, Zhizhong Tan et al.
TL;DR
HiMem uses hierarchical Episode Memory plus Note Memory with conflict-aware Memory Reconsolidation to reach 80.71% GPT-Score on LoCoMo vs 69.03% for SeCom (+11.68 points).
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
Existing long-term memory systems still show clear limitations in adaptability, scalability, and self-evolution under continuous interaction settings.
When long-horizon conversational agents lose semantic alignment, they mis-handle temporal references and coreference, degrading long-term consistency and personalization over sustained interactions.
HOW IT WORKS
HiMem’s core mechanism combines Episode Memory, Note Memory, Knowledge Alignment, and Memory Reconsolidation into a hierarchical long-term memory framework for dialogues.
You can think of HiMem like a brain with a hippocampus for detailed events and a card catalog of abstract notes, linked through alignment and reconsolidation.
This hierarchical design lets HiMem keep both concrete episodes and abstract knowledge, enabling conflict-aware updates that a plain context window cannot support.
DIAGRAM
This diagram shows how HiMem performs best-effort retrieval from Note Memory and Episode Memory and triggers Memory Reconsolidation when evidence is insufficient.
DIAGRAM
This diagram shows how HiMem is evaluated on LoCoMo with different memory configurations and metrics.
PROCESS
Memory Construction
HiMem first runs Dual Channel Segmentation to build Episode Memory and applies Multi Stage Information Extraction to populate Note Memory from dialogue logs.
Knowledge Alignment
HiMem performs Knowledge Alignment to normalize time, resolve coreference, and link Episode Memory with Note Memory in a unified semantic space.
Memory Retrieval
HiMem uses the hierarchical memory retrieval module, starting with Note First Retrieval and optionally descending to Episode Retrieval under best effort retrieval.
Memory Reconsolidation
When Note First Retrieval fails but episodes suffice, HiMem triggers Memory Reconsolidation to ADD, UPDATE, or DELETE notes based on episodic evidence.
KEY CONTRIBUTIONS
Hierarchical Long Term Memory Framework
HiMem integrates Episode Memory and Note Memory into a hierarchical structure that bridges concrete events and abstract knowledge for long horizon conversational agents.
Topic Aware Dual Channel Segmentation
HiMem introduces Topic Aware Event Surprise Dual Channel Segmentation and Multi Stage Information Extraction to build cognitively consistent Episode Memory and Note Memory.
Conflict Aware Memory Reconsolidation
HiMem designs conflict aware Memory Reconsolidation with ADD, UPDATE, DELETE operations, improving Note Memory GPT Score by about 5.85% when Memory Self Evolution is enabled.
RESULTS
Overall GPT-Score
80.71% (mean)
+11.68 over SeCom
Overall F1
34.95% (mean)
vs SeCom 16.77% overall F1
Multi Hop GPT-Score
70.92% (mean)
+11.82 over Mem0 on Multi Hop
Temporal GPT-Score
74.77% (mean)
+6.23 over Mem0 on Temporal
On the LoCoMo long horizon dialogue benchmark, which averages 600 turns and 16K tokens per session, HiMem’s 80.71% overall GPT-Score shows more reliable long term reasoning than SeCom and Mem0. These gains demonstrate that HiMem’s hierarchical memory and reconsolidation design improves multi hop and temporal reasoning under extended interactions.
BENCHMARK
On the LoCoMo long horizon dialogue benchmark, which averages 600 turns and 16K tokens per session, HiMem’s 80.71% overall GPT-Score shows more reliable long term reasoning than SeCom and Mem0. These gains demonstrate that HiMem’s hierarchical memory and reconsolidation design improves multi hop and temporal reasoning under extended interactions.
BENCHMARK
Overall GPT-Score on LoCoMo for HiMem and baseline long term memory systems.
BENCHMARK
Overall GPT-Score for HiMem and its variants without Episode Memory or Note Memory.
KEY INSIGHT
Disabling Knowledge Alignment for Episode Memory slightly improves average GPT-Score from 78.12% to 79.63%, despite alignment helping Note Memory substantially.
This is surprising because you might expect more alignment to always help, but HiMem shows that over aligning episodic context can dilute useful raw dialogue signals.
WHY IT MATTERS
HiMem unlocks long horizon agents that maintain both detailed episodic histories and compact, self evolving knowledge without sacrificing retrieval efficiency.
Builders can now deploy conversational agents that adapt their Note Memory over months of interaction, correcting contradictions instead of blindly accumulating stale or inconsistent facts.
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…