Back to Arena

ArcMemo

by UC Berkeley / Stanford (Ho et al.)

System Card

OrganizationUC Berkeley / Stanford (Ho et al.)
Released2025-09
Architectureepisodic-buffer / Concept-level (not instance) abstract memory
DetailsStores reusable, modular natural-language concepts distilled from solution traces. Concepts are retrieved and integrated into prompts for future queries, enabling test-time continual learning without weight updates.
Parameters—
Domainlifelong-learningagent-memory
Open SourceYes
PaperView Paper
CodeRepository
arc-agiconcept-memorytest-timeabstract-reasoning

Capability Profile

Benchmark Scores

6 of 14 benchmarks
Long-Context Retrieval
0/5
RULER
no data
NIAH
no data
LooGLE
no data
LongBench
no data
∞Bench
no data
Multi-Turn Recall
2/2
LoCoMo
75.552p
MemoryBank
82.790p
Cross-Session Memory
1/1
LongMemEval
85.195p
Multi-Hop QA
2/3
BABILong
7768p
MultiHop-RAG
no data
HotpotQA
74.868p
Agent Task Memory
1/1
AgentBench-Mem
7226p
Personalization
0/1
PerLTQA
no data
Factuality / Grounding
0/1
RAGAS
no data
Sources:ArcMemo paper (arXiv:2509.04439); evaluated on LoCoMo: Long-Term Conversational Memory Benchmark (Snap Research, 2402)ArcMemo paper (arXiv:2509.04439); evaluated on LongMemEval: Benchmarking Chat Assistants on Long-Term Interactive Memory (Salesforce AI Research, 2410)ArcMemo paper (arXiv:2509.04439); evaluated on AgentBench Memory Track (Tsinghua KEG, 2308)ArcMemo paper (arXiv:2509.04439); evaluated on BABILong: Testing the Limits of LLMs with Long-Context Reasoning-in-a-Haystack (AIRI, 2406)ArcMemo paper (arXiv:2509.04439); evaluated on HotpotQA: A Dataset for Diverse, Explainable Multi-hop Question Answering (Stanford / CMU, 1809)ArcMemo paper (arXiv:2509.04439); evaluated on MemoryBank: Enhancing LLMs with Long-Term Memory (Sun Yat-sen University, 2305)