Memory Agent Performance on MemoryAgentBench
83Access RateInfini Memory-A
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| Infini Memory-ABackbone=gpt-5-mini, Input Chunk Size=4096 tokens, Evaluation Protocol=LLM-as-Judge (gpt-5), Variant=Agentic Retrieval2026.06 | 83 | — | 84 | 59.9 | 81 | 64.7 | |
| HippoRAG-v2Backbone=gpt-5-mini, Input Chunk Size=4096 tokens, Evaluation Protocol=LLM-as-Judge (gpt-5)2026.06 | 78 | — | 65.2 | 32.1 | 58 | 45.5 | |
| Infini Memory-HBackbone=gpt-5-mini, Input Chunk Size=4096 tokens, Evaluation Protocol=LLM-as-Judge (gpt-5), Variant=Hybrid2026.06 | 77 | — | 81 | 57.2 | 80 | 61.3 | |
| REMemBackbone=gpt-5-mini, Input Chunk Size=4096 tokens, Evaluation Protocol=LLM-as-Judge (gpt-5)2026.06 | 62 | — | 80 | 23.4 | 32 | 38.6 | |
| LightMemBackbone=gpt-5-mini, Input Chunk Size=4096 tokens, Evaluation Protocol=LLM-as-Judge (gpt-5)2026.06 | 57 | — | 58 | 27.5 | 29 | 34.2 | |
| MemGPTBackbone=gpt-5-mini, Input Chunk Size=4096 tokens, Evaluation Protocol=LLM-as-Judge (gpt-5)2026.06 | 49 | — | 65.6 | 10.9 | 32 | 31.2 | |
| δ-MemBackbone=Qwen3-8B, Memory Configuration=TSW2026.05 | 45.65 | 32.97 | 16.43 | 46.41 | 12.88 | — | |
| δ-MemBackbone=Qwen3-8B, Memory Configuration=SSW2026.05 | 45.55 | 33.63 | 19.52 | 46.11 | 13.5 | — | |
| δ-MemBackbone=Qwen3-8B, Memory Configuration=MSW2026.05 | 45.55 | 32.66 | 14.95 | 44.6 | 13.38 | — | |
| Qwen3-8BBackbone=Qwen3-8B2026.05 | 45.1 | 31.87 | 12.79 | 48.22 | 12 | — | |
| δ-Mem (MSW)Backbone=Qwen3-4B-Instruct, Memory Mechanism=δ-Mem2026.05 | 44.4 | 38.85 | 47.29 | 41.55 | 17 | — | |
| δ-MemBackbone=Qwen3-4B-Instruct, Memory Configuration=MSW2026.05 | 44.4 | 38.85 | 47.29 | 41.55 | 17 | — | |
| δ-Mem (MSW)Writing Granularity=Multi-State2026.06 | 44.4 | 38.85 | 47.29 | 41.55 | 17 | — | |
| Context2LoRABackbone=Qwen3-8B2026.05 | 43.15 | 30.52 | 10.05 | 34.04 | 16.13 | — | |
| δ-Mem (TSW)Backbone=Qwen3-4B-Instruct, Memory Mechanism=δ-Mem2026.05 | 42.45 | 36.48 | 40.64 | 46.08 | 15.88 | — | |
| δ-MemBackbone=Qwen3-4B-Instruct, Memory Configuration=TSW2026.05 | 42.45 | 36.48 | 40.64 | 46.08 | 15.88 | — | |
| δ-Mem (TSW)Writing Granularity=Token-State2026.06 | 42.45 | 36.48 | 40.64 | 46.08 | 15.88 | — | |
| δ-Mem (SSW)Backbone=Qwen3-4B-Instruct, Memory Mechanism=δ-Mem2026.05 | 41.5 | 37.84 | 50.5 | 43.02 | 16.5 | — | |
| δ-MemBackbone=Qwen3-4B-Instruct, Memory Configuration=SSW2026.05 | 41.5 | 37.84 | 50.5 | 43.02 | 16.5 | — | |
| δ-Mem (SSW)Writing Granularity=Sequence-State2026.06 | 41.5 | 37.84 | 50.5 | 43.02 | 16.5 | — | |
| Context2LoRABackbone=Qwen3-4B-Instruct, Memory Mechanism=Parametric Memory2026.05 | 40 | 32.53 | 29.86 | 25.15 | 17.75 | — | |
| Context2LoRABackbone=Qwen3-4B-Instruct2026.05 | 40 | 32.53 | 29.86 | 25.15 | 17.75 | — | |
| Context2LoRAMemory Category=Parametric Memory2026.06 | 40 | 32.53 | 29.86 | 25.15 | 17.75 | — | |
| MLP MemoryBackbone=Qwen3-4B-Instruct, Memory Mechanism=Outside-channel Memory2026.05 | 35.35 | 28.8 | 26 | 31.19 | 14.38 | — | |
| MLP MemoryMemory Category=Outside-channel Memory2026.06 | 35.35 | 28.8 | 26 | 31.19 | 14.38 | — | |
| Qwen3-4B-InstructBackbone=Qwen3-4B-Instruct2026.05 | 35.3 | 29.54 | 26.14 | 47.08 | 14.37 | — | |
| Qwen3-4B-InstructBackbone=Qwen3-4B-Instruct2026.05 | 35.3 | 29.54 | 26.14 | 47.08 | 14.37 | — | |
| Qwen3-4B-InstructModel Type=Base Model2026.06 | 35.3 | 29.54 | 26.14 | 47.08 | 14.37 | — | |
| MemoRAGBackbone=gpt-5-mini, Input Chunk Size=4096 tokens, Evaluation Protocol=LLM-as-Judge (gpt-5)2026.06 | 35 | — | 75 | 15.3 | 25 | 33.5 | |
| MemGenBackbone=Qwen3-4B-Instruct, Memory Mechanism=Parametric Memory2026.05 | 34.85 | 29.61 | 28.45 | 44.3 | 14.38 | — | |
| MemGenMemory Category=Parametric Memory2026.06 | 34.85 | 29.61 | 28.45 | 44.3 | 14.38 | — | |
| BM25 RAGBackbone=Qwen3-4B-Instruct, Memory Mechanism=Textual Memory2026.05 | 32.2 | 24.49 | 9.74 | 37.86 | 15 | — | |
| BM25 RAGMemory Category=Textual Memory2026.06 | 32.2 | 24.49 | 9.74 | 37.86 | 15 | — | |
| RAPTORBackbone=gpt-5-mini, Input Chunk Size=4096 tokens, Evaluation Protocol=LLM-as-Judge (gpt-5)2026.06 | 32 | — | 56.8 | 25.3 | 16 | 29.8 | |
| Mem0Backbone=gpt-5-mini, Input Chunk Size=4096 tokens, Evaluation Protocol=LLM-as-Judge (gpt-5)2026.06 | 28 | — | 35.2 | 12.4 | 25 | 25 | |
| MemoryBankBackbone=Qwen3-4B-Instruct, Memory Mechanism=Textual Memory2026.05 | 22.65 | 17.65 | 7.67 | 36.36 | 9.88 | — | |
| MemoryBankMemory Category=Textual Memory2026.06 | 22.65 | 17.65 | 7.67 | 36.36 | 9.88 | — | |
| LLMLingua-2Backbone=Qwen3-4B-Instruct, Memory Mechanism=Textual Memory2026.05 | 21.45 | 15.63 | 1.43 | 38.45 | 8.62 | — | |
| LLMLingua-2Memory Category=Textual Memory2026.06 | 21.45 | 15.63 | 1.43 | 38.45 | 8.62 | — | |
| δ-MemBackbone=SmolLM3-3B, Memory Configuration=MSW2026.05 | 18.1 | 20.54 | 8.32 | 39.63 | 16.12 | — | |
| δ-MemBackbone=SmolLM3-3B, Memory Configuration=TSW2026.05 | 17.98 | 20.74 | 8.71 | 41.63 | 15.75 | — | |
| δ-MemBackbone=SmolLM3-3B, Memory Configuration=SSW2026.05 | 16.8 | 19.22 | 6.5 | 37.84 | 15.75 | — | |
| Context2LoRABackbone=SmolLM3-3B2026.05 | 16.08 | 17.62 | 2.86 | 36.77 | 14.75 | — | |
| SmolLM3-3BBackbone=SmolLM3-3B2026.05 | 12.57 | 14.21 | 5.53 | 30.72 | 8 | — |