--- title: "BEAM leaderboard: long-term memory benchmark results" description: "Every published result on BEAM, the ICLR 2026 long-term memory benchmark, at 100K, 500K, 1M and 10M tokens, each figure linked to its source." canonical: https://past.dev/benchmarks/beam last-updated: 2026-09-17 --- # BEAM leaderboard > BEAM (Beyond a Million Tokens, Tavakoli et al., ICLR 2026) measures long-term memory in AI agents. It asks 2,000 questions over 100 conversations at four history sizes: 100K, 500K, 1M and 10M tokens. past.dev ranks first at every size among published results: 91.44% at 100K, 89.41% at 500K, 88.82% at 1M and 83.97% at 10M tokens. Every figure on this page links to the result it comes from. Source: https://past.dev/benchmarks/beam ## BEAM accuracy by history size | System | 100K | 500K | 1M | 10M | Source | | --- | --- | --- | --- | --- | --- | | past.dev | 91.44% | 89.41% | 88.82% | 83.97% | [past.dev](https://past.dev/benchmarks) | | Exabase M-1 | 76.9% | not published | 75.0% | 68.0% | [exabase.io](https://exabase.io/blog/exabase-m1-achieves-state-of-the-art-on-beam-benchmark) | | Hindsight | 75.0% | 71.1% | 73.9% | 64.1% | [benchmarks.hindsight.vectorize.io](https://benchmarks.hindsight.vectorize.io/) | | Honcho | 63.0% | 64.9% | 63.1% | 40.6% | [plasticlabs.ai](https://plasticlabs.ai/blog/research/Benchmarking-Honcho) | | mem0 | not published | not published | 64.1% | 48.6% | [mem0.ai](https://mem0.ai/blog/ai-memory-benchmarks-in-2026) | - 91.44% at 100K tokens, 14.5 points above Exabase M-1, the next published result at this size. - 89.41% at 500K tokens, 18.3 points above Hindsight, the next published result at this size. - 88.82% at 1M tokens, 13.8 points above Exabase M-1, the next published result at this size. - 83.97% at 10M tokens, 16.0 points above Exabase M-1, the next published result at this size. Retrieved September 16, 2026. ## Ranked at each size ### 100K tokens 1. past.dev: 91.44% ([past.dev](https://past.dev/benchmarks)) 2. Exabase M-1: 76.9% ([exabase.io](https://exabase.io/blog/exabase-m1-achieves-state-of-the-art-on-beam-benchmark)) 3. Hindsight: 75.0% ([benchmarks.hindsight.vectorize.io](https://benchmarks.hindsight.vectorize.io/)) 4. Honcho: 63.0% ([plasticlabs.ai](https://plasticlabs.ai/blog/research/Benchmarking-Honcho)) ### 500K tokens 1. past.dev: 89.41% ([past.dev](https://past.dev/benchmarks)) 2. Hindsight: 71.1% ([benchmarks.hindsight.vectorize.io](https://benchmarks.hindsight.vectorize.io/)) 3. Honcho: 64.9% ([plasticlabs.ai](https://plasticlabs.ai/blog/research/Benchmarking-Honcho)) ### 1M tokens 1. past.dev: 88.82% ([past.dev](https://past.dev/benchmarks)) 2. Exabase M-1: 75.0% ([exabase.io](https://exabase.io/blog/exabase-m1-achieves-state-of-the-art-on-beam-benchmark)) 3. Hindsight: 73.9% ([benchmarks.hindsight.vectorize.io](https://benchmarks.hindsight.vectorize.io/)) 4. mem0: 64.1% ([mem0.ai](https://mem0.ai/blog/ai-memory-benchmarks-in-2026)) 5. Honcho: 63.1% ([plasticlabs.ai](https://plasticlabs.ai/blog/research/Benchmarking-Honcho)) ### 10M tokens 1. past.dev: 83.97% ([past.dev](https://past.dev/benchmarks)) 2. Exabase M-1: 68.0% ([exabase.io](https://exabase.io/blog/exabase-m1-achieves-state-of-the-art-on-beam-benchmark)) 3. Hindsight: 64.1% ([benchmarks.hindsight.vectorize.io](https://benchmarks.hindsight.vectorize.io/)) 4. mem0: 48.6% ([mem0.ai](https://mem0.ai/blog/ai-memory-benchmarks-in-2026)) 5. Honcho: 40.6% ([plasticlabs.ai](https://plasticlabs.ai/blog/research/Benchmarking-Honcho)) ## How to read the table Every competitor figure is that vendor's own published result, linked in the table. These are separately published runs rather than one harness, so models, judges and configurations differ between rows, and the past.dev figures are complete splits at every size. A model given the whole history in its context, with no memory system, scores under 30% on BEAM at 128K tokens and under 14% at 10M in the benchmark's own paper (Qwen 2.5, Llama 4 Maverick, Gemini 2 Flash, GPT-4.1 nano). [The BEAM paper](https://arxiv.org/abs/2510.27246) ## Results published in research papers These runs use different reader models from the product results above, so they are listed separately. | System | Size | Accuracy | Setup | Paper | | --- | --- | --- | --- | --- | | MEMIR | 100K | 48.26% | GPT-4.1-mini as the reader, average over the ten abilities, 400 questions | [Jin et al., 2026](https://arxiv.org/abs/2605.25869) | ## What BEAM measures | History size | Conversations | Questions | | --- | --- | --- | | 100K tokens | 20 | 400 | | 500K tokens | 35 | 700 | | 1M tokens | 35 | 700 | | 10M tokens | 10 | 200 | Each question tests one of ten abilities: abstention, contradiction resolution, event ordering, information extraction, instruction following, knowledge update, multi-session reasoning, preference following, summarization, temporal reasoning. The paper: [https://arxiv.org/abs/2510.27246](https://arxiv.org/abs/2510.27246). The dataset and code: [https://github.com/mohammadtavakoli78/BEAM](https://github.com/mohammadtavakoli78/BEAM). ## Another benchmark named BEAM Eywa (Joshi, 2026) reports results on a benchmark also named BEAM: a 700-question technical-memory suite introduced in that paper and scored with rubric nuggets. It is a different benchmark, and its scores do not compare with the results on this page. [Eywa on arXiv](https://arxiv.org/abs/2605.30771) ## Related - [past.dev benchmark results and the per-ability breakdown](https://past.dev/benchmarks) - [How to benchmark agent memory](https://past.dev/benchmarks/methodology) - [LongMemEval, explained](https://past.dev/benchmarks/longmemeval) - [LoCoMo, explained](https://past.dev/benchmarks/locomo) - [How memory accuracy degrades with scale](https://past.dev/blog/memory-accuracy-at-scale) - [What a good AI memory benchmark measures](https://past.dev/blog/ai-memory-benchmark-guide)