Back to Arena

HEMA

by independent (Ahn et al.)

System Card

Organizationindependent (Ahn et al.)
Released2025-04
Architecturehierarchical-summary / Dual Compact + Vector memory
DetailsCombines Compact Memory (a one-sentence narrative summary continuously updated) with Vector Memory (episodic chunk embeddings queried by cosine similarity). On a 6B transformer, maintains dialogues beyond 300 turns under 3,500 tokens prompt length.
Parameters—
Domainepisodic-sessionlong-contextagent-memory
Open SourceNo
PaperView Paper
hippocampusdual-memorycompact-summaryconversational

Capability Profile

Benchmark Scores

6 of 14 benchmarks
Long-Context Retrieval
1/5
RULER
no data
NIAH
no data
LooGLE
no data
LongBench
no data
∞Bench
77.231p
Multi-Turn Recall
1/2
LoCoMo
76.765p
MemoryBank
no data
Cross-Session Memory
1/1
LongMemEval
66.821p
Multi-Hop QA
2/3
BABILong
72.926p
MultiHop-RAG
no data
HotpotQA
66.735p
Agent Task Memory
1/1
AgentBench-Mem
6816p
Personalization
0/1
PerLTQA
no data
Factuality / Grounding
0/1
RAGAS
no data
Sources:HEMA paper (arXiv:2504.16754); evaluated on AgentBench Memory Track (Tsinghua KEG, 2308)HEMA paper (arXiv:2504.16754); evaluated on LoCoMo: Long-Term Conversational Memory Benchmark (Snap Research, 2402)HEMA paper (arXiv:2504.16754); evaluated on LongMemEval: Benchmarking Chat Assistants on Long-Term Interactive Memory (Salesforce AI Research, 2410)HEMA paper (arXiv:2504.16754); evaluated on BABILong: Testing the Limits of LLMs with Long-Context Reasoning-in-a-Haystack (AIRI, 2406)HEMA paper (arXiv:2504.16754); evaluated on HotpotQA: A Dataset for Diverse, Explainable Multi-hop Question Answering (Stanford / CMU, 1809)HEMA paper (arXiv:2504.16754); evaluated on InfiniteBench: Extending Long Context Evaluation Beyond 100K Tokens (Tsinghua / OpenBMB, 2402)