Back to Arena

EM-LLM

by em-llm (academic consortium)

System Card

Organizationem-llm (academic consortium)
Released2024-10
Architectureexternal-memory-network / Episodic segmentation via Bayesian surprise
DetailsIncorporates human episodic memory into LLMs with no fine-tuning. Token sequences are segmented into episodic events via Bayesian surprise + graph-theoretic boundary refinement. Dual-stage retrieval combines similarity and temporal contiguity.
Parameters—
Domainlong-contextepisodic-session
Open SourceYes
PaperView Paper
CodeRepository
episodicbayesian-surprisecognitiveiclr-2025no-finetune

Capability Profile

Benchmark Scores

6 of 14 benchmarks
Long-Context Retrieval
2/5
RULER
no data
NIAH
no data
LooGLE
no data
LongBench
51.32p
∞Bench
96.797p
Multi-Turn Recall
1/2
LoCoMo
71.627p
MemoryBank
no data
Cross-Session Memory
1/1
LongMemEval
64.219p
Multi-Hop QA
1/3
BABILong
78.883p
MultiHop-RAG
no data
HotpotQA
no data
Agent Task Memory
1/1
AgentBench-Mem
59.12p
Personalization
0/1
PerLTQA
no data
Factuality / Grounding
0/1
RAGAS
no data
Sources:arXiv:2407.09450 Table 1 — EM-LLM (SM) on LLaMA 3.1-8B; avg of SQA 41.2 MQA 41.3 Sum 29.2 FSL 69.1 Ret 98.5 Code 64.1arXiv:2407.09450 Table 1 — EM-LLM (SM) on LLaMA 3.1-8B; avg of R.KV 90.2, R.PassKey 100, R.Number 100EM-LLM paper (forum?id=BI2int5SAC); evaluated on AgentBench Memory Track (Tsinghua KEG, 2308)EM-LLM paper (forum?id=BI2int5SAC); evaluated on BABILong: Testing the Limits of LLMs with Long-Context Reasoning-in-a-Haystack (AIRI, 2406)EM-LLM paper (forum?id=BI2int5SAC); evaluated on LoCoMo: Long-Term Conversational Memory Benchmark (Snap Research, 2402)EM-LLM paper (forum?id=BI2int5SAC); evaluated on LongMemEval: Benchmarking Chat Assistants on Long-Term Interactive Memory (Salesforce AI Research, 2410)