Scale
- • 100 conversations, up to 10M tokens each, with 2,000 human-validated questions.
Memory benchmark
BEAM
A long-context conversational-memory benchmark with coherent conversations reaching 10M tokens and questions spanning multiple memory abilities.
Tasks
- • Long-context conversational QA
- • Memory-ability probing
- • Memory-system evaluation
What it measures
- • Recall across dialogue length
- • Multi-hop reasoning
- • Summarization
- • Instruction following
Original research
Beyond a Million Tokens: Benchmarking and Enhancing Long-Term Memory in LLMs
This is the paper that introduced BEAM alongside the LIGHT method.
arXiv:2510.27246 ↗Used in the field
Papers evaluated on BEAM
BenchmarkAgent Memory
ATANT v1.1: Positioning Continuity Evaluation Against Memory, Long-Context, and Agentic-Memory Benchmarks
Samuel Sameer Tanguturi
· 2026
ATANT v1.1 structurally analyzes seven benchmarks using the 7 v1.0 continuity properties, the 10 checkpoints, a property-coverage matrix, and the Kenotic v1.0 reference implementation. ATANT v1.1 reports 96% ATANT cumulative-scale versus 8.8% LOCOMO substring accuracy, showing that LOCOMO, LongMemEval, BEAM, MemoryBench, Zep eval, MemGPT/Letta, and RULER measure different properties from continuity.
Benchmark
Breaking the KV Cache Bottleneck: Fan Duality Model Achieves O(1) Decode Memory with Superior Associative Recall
Yasong Fan
· 2026
Fan Duality Model (FDM) uses the Fan Operator, Local-Global Cache, Freeze-Scan Training, and Holographic Reference Beam Decoding to separate wave-like compression from particle-like associative recall. On WikiText-103, Fan Duality Model (FDM) reaches 64.9 perplexity with Freeze-Scan and 62.79 with holographic decoding, while achieving 0.966 MQAR accuracy compared to Transformer at 0.606.
Long-Term Memory
Transferable End-to-End Optimization for Indirect Long-Term Memory Poisoning in LLM Agents
Chuanchao Zang, Jianing Wang et al.
arXiv 2026 · 2026
PIPEPOISON combines stage-level signals, chain-structured losses, stability-calibrated configuration weights, and weighted stage losses to optimize poisoning content across diverse shadow pipelines. On 12 matched memory–agent configurations, PIPEPOISON reaches 73.4% Attack Utilization Rate on LongMemEval, LoCoMo, and BEAM, beating the best baseline (MemMorph) by 19.1 percentage points.
Long-Term Memory
Wontopos Tablet 2: Measuring Multilingual and Multimodal Memory Retrieval Without Lexical Matching
Sunwoo Kim
arXiv 2026 · 2026
Wontopos Tablet 2 measures how a retrieval engine interacts with components like LongMemEval-S, BEAM-1M, Crossmodal-3600, and the re-ask mechanism without relying on lexical matching. It reports controlled multilingual and multimodal results, including a negative finding that adding captions can reduce cross-lingual photograph retrieval quality compared to pure visual matching.