Examenos

55.49% · raw 55.49 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Tencent's released post-RL BF16 memory writer comparison, arXiv 2610.22196v1. Fixed deterministic 20% subsamples; only the memory writer changes, with fixed harness version/prompts, retrieval, answer model and judge. Each writer starts from empty memory; decoding matched where APIs permit. No failed platform judgments silently counted zero. LoCoMo Judge, LongMemEval, PersonaMem-v2, Travel SPS, Shopping PS and CMB have distinct per-harness identities; CMB is an equal mean of two repository-QA evaluations. These are vendor-run memory-pipeline tests, not independent general chat benchmarks. EverMemOS held out from OPD/RL; Mem0, MemOS and Hy-Memory used in training. No controlled ablation/repeated-seed evidence. Tiny 0.01–0.02 point mean margins are explicitly not statistically resolved. Reasoning effort, exact execution dates and sample counts per 20% subset are not published. Card Table 1 / report Table 6. Four-harness macro-average. Source model: Hy-Memory-30B-A3B. DeepSeek label explicitly means the original V4.1 Flash, never 1017.

59.54% · raw 59.54 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Tencent's released post-RL BF16 memory writer comparison, arXiv 2610.22196v1. Fixed deterministic 20% subsamples; only the memory writer changes, with fixed harness version/prompts, retrieval, answer model and judge. Each writer starts from empty memory; decoding matched where APIs permit. No failed platform judgments silently counted zero. LoCoMo Judge, LongMemEval, PersonaMem-v2, Travel SPS, Shopping PS and CMB have distinct per-harness identities; CMB is an equal mean of two repository-QA evaluations. These are vendor-run memory-pipeline tests, not independent general chat benchmarks. EverMemOS held out from OPD/RL; Mem0, MemOS and Hy-Memory used in training. No controlled ablation/repeated-seed evidence. Tiny 0.01–0.02 point mean margins are explicitly not statistically resolved. Reasoning effort, exact execution dates and sample counts per 20% subset are not published. Harness: Mem0; metric: LoCoMo Judge; card Table 2 / report Table 7; exact source row: Hy-Memory-30B-A3B.

73.00% · raw 73 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Tencent's released post-RL BF16 memory writer comparison, arXiv 2610.22196v1. Fixed deterministic 20% subsamples; only the memory writer changes, with fixed harness version/prompts, retrieval, answer model and judge. Each writer starts from empty memory; decoding matched where APIs permit. No failed platform judgments silently counted zero. LoCoMo Judge, LongMemEval, PersonaMem-v2, Travel SPS, Shopping PS and CMB have distinct per-harness identities; CMB is an equal mean of two repository-QA evaluations. These are vendor-run memory-pipeline tests, not independent general chat benchmarks. EverMemOS held out from OPD/RL; Mem0, MemOS and Hy-Memory used in training. No controlled ablation/repeated-seed evidence. Tiny 0.01–0.02 point mean margins are explicitly not statistically resolved. Reasoning effort, exact execution dates and sample counts per 20% subset are not published. Harness: Mem0; metric: LongMemEval; card Table 2 / report Table 7; exact source row: Hy-Memory-30B-A3B.

51.62% · raw 51.62 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Tencent's released post-RL BF16 memory writer comparison, arXiv 2610.22196v1. Fixed deterministic 20% subsamples; only the memory writer changes, with fixed harness version/prompts, retrieval, answer model and judge. Each writer starts from empty memory; decoding matched where APIs permit. No failed platform judgments silently counted zero. LoCoMo Judge, LongMemEval, PersonaMem-v2, Travel SPS, Shopping PS and CMB have distinct per-harness identities; CMB is an equal mean of two repository-QA evaluations. These are vendor-run memory-pipeline tests, not independent general chat benchmarks. EverMemOS held out from OPD/RL; Mem0, MemOS and Hy-Memory used in training. No controlled ablation/repeated-seed evidence. Tiny 0.01–0.02 point mean margins are explicitly not statistically resolved. Reasoning effort, exact execution dates and sample counts per 20% subset are not published. Harness: Mem0; metric: PersonaMem-v2; card Table 2 / report Table 7; exact source row: Hy-Memory-30B-A3B.

54.51% · raw 54.51 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Tencent's released post-RL BF16 memory writer comparison, arXiv 2610.22196v1. Fixed deterministic 20% subsamples; only the memory writer changes, with fixed harness version/prompts, retrieval, answer model and judge. Each writer starts from empty memory; decoding matched where APIs permit. No failed platform judgments silently counted zero. LoCoMo Judge, LongMemEval, PersonaMem-v2, Travel SPS, Shopping PS and CMB have distinct per-harness identities; CMB is an equal mean of two repository-QA evaluations. These are vendor-run memory-pipeline tests, not independent general chat benchmarks. EverMemOS held out from OPD/RL; Mem0, MemOS and Hy-Memory used in training. No controlled ablation/repeated-seed evidence. Tiny 0.01–0.02 point mean margins are explicitly not statistically resolved. Reasoning effort, exact execution dates and sample counts per 20% subset are not published. Harness: Mem0; metric: Travel SPS; card Table 2 / report Table 7; exact source row: Hy-Memory-30B-A3B.

18.33% · raw 18.33 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Tencent's released post-RL BF16 memory writer comparison, arXiv 2610.22196v1. Fixed deterministic 20% subsamples; only the memory writer changes, with fixed harness version/prompts, retrieval, answer model and judge. Each writer starts from empty memory; decoding matched where APIs permit. No failed platform judgments silently counted zero. LoCoMo Judge, LongMemEval, PersonaMem-v2, Travel SPS, Shopping PS and CMB have distinct per-harness identities; CMB is an equal mean of two repository-QA evaluations. These are vendor-run memory-pipeline tests, not independent general chat benchmarks. EverMemOS held out from OPD/RL; Mem0, MemOS and Hy-Memory used in training. No controlled ablation/repeated-seed evidence. Tiny 0.01–0.02 point mean margins are explicitly not statistically resolved. Reasoning effort, exact execution dates and sample counts per 20% subset are not published. Harness: Mem0; metric: Shopping PS; card Table 2 / report Table 7; exact source row: Hy-Memory-30B-A3B.

67.13% · raw 67.13 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Tencent's released post-RL BF16 memory writer comparison, arXiv 2610.22196v1. Fixed deterministic 20% subsamples; only the memory writer changes, with fixed harness version/prompts, retrieval, answer model and judge. Each writer starts from empty memory; decoding matched where APIs permit. No failed platform judgments silently counted zero. LoCoMo Judge, LongMemEval, PersonaMem-v2, Travel SPS, Shopping PS and CMB have distinct per-harness identities; CMB is an equal mean of two repository-QA evaluations. These are vendor-run memory-pipeline tests, not independent general chat benchmarks. EverMemOS held out from OPD/RL; Mem0, MemOS and Hy-Memory used in training. No controlled ablation/repeated-seed evidence. Tiny 0.01–0.02 point mean margins are explicitly not statistically resolved. Reasoning effort, exact execution dates and sample counts per 20% subset are not published. Harness: Mem0; metric: CMB (two repository-QA mean); card Table 2 / report Table 7; exact source row: Hy-Memory-30B-A3B.

54.02% · raw 54.02 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Tencent's released post-RL BF16 memory writer comparison, arXiv 2610.22196v1. Fixed deterministic 20% subsamples; only the memory writer changes, with fixed harness version/prompts, retrieval, answer model and judge. Each writer starts from empty memory; decoding matched where APIs permit. No failed platform judgments silently counted zero. LoCoMo Judge, LongMemEval, PersonaMem-v2, Travel SPS, Shopping PS and CMB have distinct per-harness identities; CMB is an equal mean of two repository-QA evaluations. These are vendor-run memory-pipeline tests, not independent general chat benchmarks. EverMemOS held out from OPD/RL; Mem0, MemOS and Hy-Memory used in training. No controlled ablation/repeated-seed evidence. Tiny 0.01–0.02 point mean margins are explicitly not statistically resolved. Reasoning effort, exact execution dates and sample counts per 20% subset are not published. Harness: Mem0; metric: Six-metric mean; card Table 2 / report Table 7; exact source row: Hy-Memory-30B-A3B.

73.28% · raw 73.28 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Tencent's released post-RL BF16 memory writer comparison, arXiv 2610.22196v1. Fixed deterministic 20% subsamples; only the memory writer changes, with fixed harness version/prompts, retrieval, answer model and judge. Each writer starts from empty memory; decoding matched where APIs permit. No failed platform judgments silently counted zero. LoCoMo Judge, LongMemEval, PersonaMem-v2, Travel SPS, Shopping PS and CMB have distinct per-harness identities; CMB is an equal mean of two repository-QA evaluations. These are vendor-run memory-pipeline tests, not independent general chat benchmarks. EverMemOS held out from OPD/RL; Mem0, MemOS and Hy-Memory used in training. No controlled ablation/repeated-seed evidence. Tiny 0.01–0.02 point mean margins are explicitly not statistically resolved. Reasoning effort, exact execution dates and sample counts per 20% subset are not published. Harness: Hy-Memory; metric: LoCoMo Judge; card Table 3 / report Table 8; exact source row: Hy-Memory-30B-A3B.

89.00% · raw 89 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Tencent's released post-RL BF16 memory writer comparison, arXiv 2610.22196v1. Fixed deterministic 20% subsamples; only the memory writer changes, with fixed harness version/prompts, retrieval, answer model and judge. Each writer starts from empty memory; decoding matched where APIs permit. No failed platform judgments silently counted zero. LoCoMo Judge, LongMemEval, PersonaMem-v2, Travel SPS, Shopping PS and CMB have distinct per-harness identities; CMB is an equal mean of two repository-QA evaluations. These are vendor-run memory-pipeline tests, not independent general chat benchmarks. EverMemOS held out from OPD/RL; Mem0, MemOS and Hy-Memory used in training. No controlled ablation/repeated-seed evidence. Tiny 0.01–0.02 point mean margins are explicitly not statistically resolved. Reasoning effort, exact execution dates and sample counts per 20% subset are not published. Harness: Hy-Memory; metric: LongMemEval; card Table 3 / report Table 8; exact source row: Hy-Memory-30B-A3B.

51.54% · raw 51.54 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Tencent's released post-RL BF16 memory writer comparison, arXiv 2610.22196v1. Fixed deterministic 20% subsamples; only the memory writer changes, with fixed harness version/prompts, retrieval, answer model and judge. Each writer starts from empty memory; decoding matched where APIs permit. No failed platform judgments silently counted zero. LoCoMo Judge, LongMemEval, PersonaMem-v2, Travel SPS, Shopping PS and CMB have distinct per-harness identities; CMB is an equal mean of two repository-QA evaluations. These are vendor-run memory-pipeline tests, not independent general chat benchmarks. EverMemOS held out from OPD/RL; Mem0, MemOS and Hy-Memory used in training. No controlled ablation/repeated-seed evidence. Tiny 0.01–0.02 point mean margins are explicitly not statistically resolved. Reasoning effort, exact execution dates and sample counts per 20% subset are not published. Harness: Hy-Memory; metric: PersonaMem-v2; card Table 3 / report Table 8; exact source row: Hy-Memory-30B-A3B.

57.65% · raw 57.65 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Tencent's released post-RL BF16 memory writer comparison, arXiv 2610.22196v1. Fixed deterministic 20% subsamples; only the memory writer changes, with fixed harness version/prompts, retrieval, answer model and judge. Each writer starts from empty memory; decoding matched where APIs permit. No failed platform judgments silently counted zero. LoCoMo Judge, LongMemEval, PersonaMem-v2, Travel SPS, Shopping PS and CMB have distinct per-harness identities; CMB is an equal mean of two repository-QA evaluations. These are vendor-run memory-pipeline tests, not independent general chat benchmarks. EverMemOS held out from OPD/RL; Mem0, MemOS and Hy-Memory used in training. No controlled ablation/repeated-seed evidence. Tiny 0.01–0.02 point mean margins are explicitly not statistically resolved. Reasoning effort, exact execution dates and sample counts per 20% subset are not published. Harness: Hy-Memory; metric: Travel SPS; card Table 3 / report Table 8; exact source row: Hy-Memory-30B-A3B.

21.11% · raw 21.11 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Tencent's released post-RL BF16 memory writer comparison, arXiv 2610.22196v1. Fixed deterministic 20% subsamples; only the memory writer changes, with fixed harness version/prompts, retrieval, answer model and judge. Each writer starts from empty memory; decoding matched where APIs permit. No failed platform judgments silently counted zero. LoCoMo Judge, LongMemEval, PersonaMem-v2, Travel SPS, Shopping PS and CMB have distinct per-harness identities; CMB is an equal mean of two repository-QA evaluations. These are vendor-run memory-pipeline tests, not independent general chat benchmarks. EverMemOS held out from OPD/RL; Mem0, MemOS and Hy-Memory used in training. No controlled ablation/repeated-seed evidence. Tiny 0.01–0.02 point mean margins are explicitly not statistically resolved. Reasoning effort, exact execution dates and sample counts per 20% subset are not published. Harness: Hy-Memory; metric: Shopping PS; card Table 3 / report Table 8; exact source row: Hy-Memory-30B-A3B.

70.73% · raw 70.73 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Tencent's released post-RL BF16 memory writer comparison, arXiv 2610.22196v1. Fixed deterministic 20% subsamples; only the memory writer changes, with fixed harness version/prompts, retrieval, answer model and judge. Each writer starts from empty memory; decoding matched where APIs permit. No failed platform judgments silently counted zero. LoCoMo Judge, LongMemEval, PersonaMem-v2, Travel SPS, Shopping PS and CMB have distinct per-harness identities; CMB is an equal mean of two repository-QA evaluations. These are vendor-run memory-pipeline tests, not independent general chat benchmarks. EverMemOS held out from OPD/RL; Mem0, MemOS and Hy-Memory used in training. No controlled ablation/repeated-seed evidence. Tiny 0.01–0.02 point mean margins are explicitly not statistically resolved. Reasoning effort, exact execution dates and sample counts per 20% subset are not published. Harness: Hy-Memory; metric: CMB (two repository-QA mean); card Table 3 / report Table 8; exact source row: Hy-Memory-30B-A3B.

60.55% · raw 60.55 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Tencent's released post-RL BF16 memory writer comparison, arXiv 2610.22196v1. Fixed deterministic 20% subsamples; only the memory writer changes, with fixed harness version/prompts, retrieval, answer model and judge. Each writer starts from empty memory; decoding matched where APIs permit. No failed platform judgments silently counted zero. LoCoMo Judge, LongMemEval, PersonaMem-v2, Travel SPS, Shopping PS and CMB have distinct per-harness identities; CMB is an equal mean of two repository-QA evaluations. These are vendor-run memory-pipeline tests, not independent general chat benchmarks. EverMemOS held out from OPD/RL; Mem0, MemOS and Hy-Memory used in training. No controlled ablation/repeated-seed evidence. Tiny 0.01–0.02 point mean margins are explicitly not statistically resolved. Reasoning effort, exact execution dates and sample counts per 20% subset are not published. Harness: Hy-Memory; metric: Six-metric mean; card Table 3 / report Table 8; exact source row: Hy-Memory-30B-A3B.

62.98% · raw 62.98 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Tencent's released post-RL BF16 memory writer comparison, arXiv 2610.22196v1. Fixed deterministic 20% subsamples; only the memory writer changes, with fixed harness version/prompts, retrieval, answer model and judge. Each writer starts from empty memory; decoding matched where APIs permit. No failed platform judgments silently counted zero. LoCoMo Judge, LongMemEval, PersonaMem-v2, Travel SPS, Shopping PS and CMB have distinct per-harness identities; CMB is an equal mean of two repository-QA evaluations. These are vendor-run memory-pipeline tests, not independent general chat benchmarks. EverMemOS held out from OPD/RL; Mem0, MemOS and Hy-Memory used in training. No controlled ablation/repeated-seed evidence. Tiny 0.01–0.02 point mean margins are explicitly not statistically resolved. Reasoning effort, exact execution dates and sample counts per 20% subset are not published. Harness: MemOS; metric: LoCoMo Judge; card Table 4 / report Table 9; exact source row: Hy-Memory-30B-A3B.

73.00% · raw 73 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Tencent's released post-RL BF16 memory writer comparison, arXiv 2610.22196v1. Fixed deterministic 20% subsamples; only the memory writer changes, with fixed harness version/prompts, retrieval, answer model and judge. Each writer starts from empty memory; decoding matched where APIs permit. No failed platform judgments silently counted zero. LoCoMo Judge, LongMemEval, PersonaMem-v2, Travel SPS, Shopping PS and CMB have distinct per-harness identities; CMB is an equal mean of two repository-QA evaluations. These are vendor-run memory-pipeline tests, not independent general chat benchmarks. EverMemOS held out from OPD/RL; Mem0, MemOS and Hy-Memory used in training. No controlled ablation/repeated-seed evidence. Tiny 0.01–0.02 point mean margins are explicitly not statistically resolved. Reasoning effort, exact execution dates and sample counts per 20% subset are not published. Harness: MemOS; metric: LongMemEval; card Table 4 / report Table 9; exact source row: Hy-Memory-30B-A3B.

51.92% · raw 51.92 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Tencent's released post-RL BF16 memory writer comparison, arXiv 2610.22196v1. Fixed deterministic 20% subsamples; only the memory writer changes, with fixed harness version/prompts, retrieval, answer model and judge. Each writer starts from empty memory; decoding matched where APIs permit. No failed platform judgments silently counted zero. LoCoMo Judge, LongMemEval, PersonaMem-v2, Travel SPS, Shopping PS and CMB have distinct per-harness identities; CMB is an equal mean of two repository-QA evaluations. These are vendor-run memory-pipeline tests, not independent general chat benchmarks. EverMemOS held out from OPD/RL; Mem0, MemOS and Hy-Memory used in training. No controlled ablation/repeated-seed evidence. Tiny 0.01–0.02 point mean margins are explicitly not statistically resolved. Reasoning effort, exact execution dates and sample counts per 20% subset are not published. Harness: MemOS; metric: PersonaMem-v2; card Table 4 / report Table 9; exact source row: Hy-Memory-30B-A3B.

24.15% · raw 24.15 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Tencent's released post-RL BF16 memory writer comparison, arXiv 2610.22196v1. Fixed deterministic 20% subsamples; only the memory writer changes, with fixed harness version/prompts, retrieval, answer model and judge. Each writer starts from empty memory; decoding matched where APIs permit. No failed platform judgments silently counted zero. LoCoMo Judge, LongMemEval, PersonaMem-v2, Travel SPS, Shopping PS and CMB have distinct per-harness identities; CMB is an equal mean of two repository-QA evaluations. These are vendor-run memory-pipeline tests, not independent general chat benchmarks. EverMemOS held out from OPD/RL; Mem0, MemOS and Hy-Memory used in training. No controlled ablation/repeated-seed evidence. Tiny 0.01–0.02 point mean margins are explicitly not statistically resolved. Reasoning effort, exact execution dates and sample counts per 20% subset are not published. Harness: MemOS; metric: Travel SPS; card Table 4 / report Table 9; exact source row: Hy-Memory-30B-A3B.

21.11% · raw 21.11 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Tencent's released post-RL BF16 memory writer comparison, arXiv 2610.22196v1. Fixed deterministic 20% subsamples; only the memory writer changes, with fixed harness version/prompts, retrieval, answer model and judge. Each writer starts from empty memory; decoding matched where APIs permit. No failed platform judgments silently counted zero. LoCoMo Judge, LongMemEval, PersonaMem-v2, Travel SPS, Shopping PS and CMB have distinct per-harness identities; CMB is an equal mean of two repository-QA evaluations. These are vendor-run memory-pipeline tests, not independent general chat benchmarks. EverMemOS held out from OPD/RL; Mem0, MemOS and Hy-Memory used in training. No controlled ablation/repeated-seed evidence. Tiny 0.01–0.02 point mean margins are explicitly not statistically resolved. Reasoning effort, exact execution dates and sample counts per 20% subset are not published. Harness: MemOS; metric: Shopping PS; card Table 4 / report Table 9; exact source row: Hy-Memory-30B-A3B.

69.91% · raw 69.91 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Tencent's released post-RL BF16 memory writer comparison, arXiv 2610.22196v1. Fixed deterministic 20% subsamples; only the memory writer changes, with fixed harness version/prompts, retrieval, answer model and judge. Each writer starts from empty memory; decoding matched where APIs permit. No failed platform judgments silently counted zero. LoCoMo Judge, LongMemEval, PersonaMem-v2, Travel SPS, Shopping PS and CMB have distinct per-harness identities; CMB is an equal mean of two repository-QA evaluations. These are vendor-run memory-pipeline tests, not independent general chat benchmarks. EverMemOS held out from OPD/RL; Mem0, MemOS and Hy-Memory used in training. No controlled ablation/repeated-seed evidence. Tiny 0.01–0.02 point mean margins are explicitly not statistically resolved. Reasoning effort, exact execution dates and sample counts per 20% subset are not published. Harness: MemOS; metric: CMB (two repository-QA mean); card Table 4 / report Table 9; exact source row: Hy-Memory-30B-A3B.

50.51% · raw 50.51 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Tencent's released post-RL BF16 memory writer comparison, arXiv 2610.22196v1. Fixed deterministic 20% subsamples; only the memory writer changes, with fixed harness version/prompts, retrieval, answer model and judge. Each writer starts from empty memory; decoding matched where APIs permit. No failed platform judgments silently counted zero. LoCoMo Judge, LongMemEval, PersonaMem-v2, Travel SPS, Shopping PS and CMB have distinct per-harness identities; CMB is an equal mean of two repository-QA evaluations. These are vendor-run memory-pipeline tests, not independent general chat benchmarks. EverMemOS held out from OPD/RL; Mem0, MemOS and Hy-Memory used in training. No controlled ablation/repeated-seed evidence. Tiny 0.01–0.02 point mean margins are explicitly not statistically resolved. Reasoning effort, exact execution dates and sample counts per 20% subset are not published. Harness: MemOS; metric: Six-metric mean; card Table 4 / report Table 9; exact source row: Hy-Memory-30B-A3B.

73.28% · raw 73.28 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Tencent's released post-RL BF16 memory writer comparison, arXiv 2610.22196v1. Fixed deterministic 20% subsamples; only the memory writer changes, with fixed harness version/prompts, retrieval, answer model and judge. Each writer starts from empty memory; decoding matched where APIs permit. No failed platform judgments silently counted zero. LoCoMo Judge, LongMemEval, PersonaMem-v2, Travel SPS, Shopping PS and CMB have distinct per-harness identities; CMB is an equal mean of two repository-QA evaluations. These are vendor-run memory-pipeline tests, not independent general chat benchmarks. EverMemOS held out from OPD/RL; Mem0, MemOS and Hy-Memory used in training. No controlled ablation/repeated-seed evidence. Tiny 0.01–0.02 point mean margins are explicitly not statistically resolved. Reasoning effort, exact execution dates and sample counts per 20% subset are not published. Harness: EverMemOS held out; metric: LoCoMo Judge; card Table 5 / report Table 10; exact source row: Hy-Memory-30B-A3B.

76.00% · raw 76 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Tencent's released post-RL BF16 memory writer comparison, arXiv 2610.22196v1. Fixed deterministic 20% subsamples; only the memory writer changes, with fixed harness version/prompts, retrieval, answer model and judge. Each writer starts from empty memory; decoding matched where APIs permit. No failed platform judgments silently counted zero. LoCoMo Judge, LongMemEval, PersonaMem-v2, Travel SPS, Shopping PS and CMB have distinct per-harness identities; CMB is an equal mean of two repository-QA evaluations. These are vendor-run memory-pipeline tests, not independent general chat benchmarks. EverMemOS held out from OPD/RL; Mem0, MemOS and Hy-Memory used in training. No controlled ablation/repeated-seed evidence. Tiny 0.01–0.02 point mean margins are explicitly not statistically resolved. Reasoning effort, exact execution dates and sample counts per 20% subset are not published. Harness: EverMemOS held out; metric: LongMemEval; card Table 5 / report Table 10; exact source row: Hy-Memory-30B-A3B.

53.28% · raw 53.28 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Tencent's released post-RL BF16 memory writer comparison, arXiv 2610.22196v1. Fixed deterministic 20% subsamples; only the memory writer changes, with fixed harness version/prompts, retrieval, answer model and judge. Each writer starts from empty memory; decoding matched where APIs permit. No failed platform judgments silently counted zero. LoCoMo Judge, LongMemEval, PersonaMem-v2, Travel SPS, Shopping PS and CMB have distinct per-harness identities; CMB is an equal mean of two repository-QA evaluations. These are vendor-run memory-pipeline tests, not independent general chat benchmarks. EverMemOS held out from OPD/RL; Mem0, MemOS and Hy-Memory used in training. No controlled ablation/repeated-seed evidence. Tiny 0.01–0.02 point mean margins are explicitly not statistically resolved. Reasoning effort, exact execution dates and sample counts per 20% subset are not published. Harness: EverMemOS held out; metric: PersonaMem-v2; card Table 5 / report Table 10; exact source row: Hy-Memory-30B-A3B.

56.56% · raw 56.56 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Tencent's released post-RL BF16 memory writer comparison, arXiv 2610.22196v1. Fixed deterministic 20% subsamples; only the memory writer changes, with fixed harness version/prompts, retrieval, answer model and judge. Each writer starts from empty memory; decoding matched where APIs permit. No failed platform judgments silently counted zero. LoCoMo Judge, LongMemEval, PersonaMem-v2, Travel SPS, Shopping PS and CMB have distinct per-harness identities; CMB is an equal mean of two repository-QA evaluations. These are vendor-run memory-pipeline tests, not independent general chat benchmarks. EverMemOS held out from OPD/RL; Mem0, MemOS and Hy-Memory used in training. No controlled ablation/repeated-seed evidence. Tiny 0.01–0.02 point mean margins are explicitly not statistically resolved. Reasoning effort, exact execution dates and sample counts per 20% subset are not published. Harness: EverMemOS held out; metric: Travel SPS; card Table 5 / report Table 10; exact source row: Hy-Memory-30B-A3B.

19.05% · raw 19.05 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Tencent's released post-RL BF16 memory writer comparison, arXiv 2610.22196v1. Fixed deterministic 20% subsamples; only the memory writer changes, with fixed harness version/prompts, retrieval, answer model and judge. Each writer starts from empty memory; decoding matched where APIs permit. No failed platform judgments silently counted zero. LoCoMo Judge, LongMemEval, PersonaMem-v2, Travel SPS, Shopping PS and CMB have distinct per-harness identities; CMB is an equal mean of two repository-QA evaluations. These are vendor-run memory-pipeline tests, not independent general chat benchmarks. EverMemOS held out from OPD/RL; Mem0, MemOS and Hy-Memory used in training. No controlled ablation/repeated-seed evidence. Tiny 0.01–0.02 point mean margins are explicitly not statistically resolved. Reasoning effort, exact execution dates and sample counts per 20% subset are not published. Harness: EverMemOS held out; metric: Shopping PS; card Table 5 / report Table 10; exact source row: Hy-Memory-30B-A3B.

63.19% · raw 63.19 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Tencent's released post-RL BF16 memory writer comparison, arXiv 2610.22196v1. Fixed deterministic 20% subsamples; only the memory writer changes, with fixed harness version/prompts, retrieval, answer model and judge. Each writer starts from empty memory; decoding matched where APIs permit. No failed platform judgments silently counted zero. LoCoMo Judge, LongMemEval, PersonaMem-v2, Travel SPS, Shopping PS and CMB have distinct per-harness identities; CMB is an equal mean of two repository-QA evaluations. These are vendor-run memory-pipeline tests, not independent general chat benchmarks. EverMemOS held out from OPD/RL; Mem0, MemOS and Hy-Memory used in training. No controlled ablation/repeated-seed evidence. Tiny 0.01–0.02 point mean margins are explicitly not statistically resolved. Reasoning effort, exact execution dates and sample counts per 20% subset are not published. Harness: EverMemOS held out; metric: CMB (two repository-QA mean); card Table 5 / report Table 10; exact source row: Hy-Memory-30B-A3B.

56.89% · raw 56.89 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Tencent's released post-RL BF16 memory writer comparison, arXiv 2610.22196v1. Fixed deterministic 20% subsamples; only the memory writer changes, with fixed harness version/prompts, retrieval, answer model and judge. Each writer starts from empty memory; decoding matched where APIs permit. No failed platform judgments silently counted zero. LoCoMo Judge, LongMemEval, PersonaMem-v2, Travel SPS, Shopping PS and CMB have distinct per-harness identities; CMB is an equal mean of two repository-QA evaluations. These are vendor-run memory-pipeline tests, not independent general chat benchmarks. EverMemOS held out from OPD/RL; Mem0, MemOS and Hy-Memory used in training. No controlled ablation/repeated-seed evidence. Tiny 0.01–0.02 point mean margins are explicitly not statistically resolved. Reasoning effort, exact execution dates and sample counts per 20% subset are not published. Harness: EverMemOS held out; metric: Six-metric mean; card Table 5 / report Table 10; exact source row: Hy-Memory-30B-A3B.