tencent/Hy-Memory-30B-A3B · Hugging Face

Source: https://huggingface.co/tencent/Hy-Memory-30B-A3B

https://huggingface.co/tencent/Hy-Memory-30B-A3B#model-introductionModel Introduction

Hy-Memory-30B-A3Bis a compact long-term-memory writer model (~3B active parameters) developed by Tencent Hunyuan to execute the write path inside external memory frameworks. It directly replaces the general-purpose LLM insideMem0,MemOS, andHy-Memory(covering Chat, Working, and Coding scenarios), deciding what information to retain, how precisely to represent it, and how new evidence should update the persistent memory bank. The model is trained in two stages—Memory On-Policy Distillation (OPD) followed by Attributed Memory Reinforcement Learning—and transfers directly to unseen memory frameworks such asEverMemOS.

Cross-harness macro-average scores on the fixed 20% benchmark matrix. Each harness score averages six metrics equally; Overall averages the four harnesses equally. EverMemOS is held out from training.

https://huggingface.co/tencent/Hy-Memory-30B-A3B#two-stage-trainingTwo-Stage Training

Figure 1: Hy-Memory-30B-A3B is trained inside the same deployed harness configurations in both stages. OPD supplies dense teacher supervision on student-written state chains; RL examines the resulting banks, assigns chunk-level credit, and carries the selected bank across windows. In the rollout grid, each cell is the bank after chunk t of rollout i, shaded by its chunk score s_i[t]. Advantages are normalized within each column (dashed: one chunk-level group), so the best rollout can differ from chunk to chunk. The outlined cell is the final bank with the highest sum_t s_i[t], which becomes the parent of the next window.

https://huggingface.co/tencent/Hy-Memory-30B-A3B#experimentsExperiments

https://huggingface.co/tencent/Hy-Memory-30B-A3B#evaluation-principlesEvaluation Principles

The evaluation changes only the memory writer. For each comparison, the harness version and prompts are fixed, retrieval configuration is fixed, and the answer model, judge, and other framework-side components are held fixed. Throughout this report, every reference to a DeepSeek Flash endpoint denotesDeepSeek V4.1 Flash. Decoding settings are matched where APIs permit, each writer builds its own bank from an empty state, and test examples and scoring rules are identical.

For computational efficiency, every benchmark in the main four-harness matrix is evaluated on a fixed 20% subsample. Each harness score is the equal-weight average of six metrics:LoCoMo Judge,LongMemEval (LME),PersonaMem-v2 (PM-v2),Travel SPS,Shopping PS, andCode Memory Bench (CMB). CMB is itself the equal-weight average of two repository QA evaluations. The overall score is the equal-weight average of the four harness scores.

https://huggingface.co/tencent/Hy-Memory-30B-A3B#final-cross-harness-resultsFinal Cross-Harness Results

Memory writerMem0Hy-Mem.MemOSEverMem.Avg.Hy-Memory-30B-A3B54.0260.5550.5156.89****55.49Claude Opus 553.4360.5350.5054.0454.62DeepSeek V4.1 Flash49.7360.2049.1755.4753.64Hy345.9958.0650.1952.3051.63Qwen3.6-35B-A3B47.5155.8747.3650.7550.37Hy3-A3B Base40.2050.3949.0550.9847.66Qwen3-30B-A3B39.8145.6745.4446.5744.37 Table 1: Six-metric means on fixed 20% benchmark subsamples. The macro average weights the four harnesses equally.

https://huggingface.co/tencent/Hy-Memory-30B-A3B#detailed-benchmark-resultsDetailed Benchmark Results

Tables 2–5 report the six benchmark scores used to compute each harness mean in Table 1 on the same fixed 20% benchmark subsamples.

https://huggingface.co/tencent/Hy-Memory-30B-A3B#table-2-detailed-results-on-the-mem0-harnessTable 2: Detailed results on the Mem0 harness

WriterLoCoMoLMEPM-v2TravelShop.CMBAvg.Hy-Memory-30B-A3B59.5473.0051.6254.5118.3367.1354.02Claude Opus 569.4766.0053.5842.9817.7870.7453.43DeepSeek V4.1 Flash58.7869.3949.5635.0716.6768.9249.73Hy341.2262.0044.4446.8717.2264.1645.99Qwen3.6-35B-A3B50.0069.0049.3836.3920.0060.3047.51Hy3-A3B Base50.0047.0040.8237.9114.4451.0340.20Qwen3-30B-A3B46.1856.0042.3231.9816.6745.6839.81

https://huggingface.co/tencent/Hy-Memory-30B-A3B#table-3-detailed-results-on-the-hy-memory-harnessTable 3: Detailed results on the Hy-Memory harness

WriterLoCoMoLMEPM-v2TravelShop.CMBAvg.Hy-Memory-30B-A3B73.2889.0051.5457.6521.1170.7360.55Claude Opus 574.43****89.6948.7151.3423.3375.7060.53DeepSeek V4.1 Flash72.9085.8649.0053.4622.7877.1860.20Hy369.4789.0052.7649.0121.6766.4458.06Qwen3.6-35B-A3B70.9984.8543.5649.8220.0065.9855.87Hy3-A3B Base66.4187.0049.8431.3617.7849.9750.39Qwen3-30B-A3B64.1282.0041.0632.8318.8935.1045.67

https://huggingface.co/tencent/Hy-Memory-30B-A3B#table-4-detailed-results-on-the-memos-harnessTable 4: Detailed results on the MemOS harness

WriterLoCoMoLMEPM-v2TravelShop.CMBAvg.Hy-Memory-30B-A3B62.9873.0051.9224.1521.1169.9150.51Claude Opus 563.7479.3852.8422.1419.4465.4650.50DeepSeek V4.1 Flash61.0773.0050.6818.6121.67****69.9749.17Hy362.9881.0049.6418.8718.8969.7850.19Qwen3.6-35B-A3B55.3479.0051.3420.0815.5662.8247.36Hy3-A3B Base58.4073.0050.8626.0521.1164.8849.05Qwen3-30B-A3B59.1673.0048.2823.0316.6752.4945.44

https://huggingface.co/tencent/Hy-Memory-30B-A3B#table-5-detailed-results-on-the-held-out-evermemos-harnessTable 5: Detailed results on the held-out EverMemOS harness

WriterLoCoMoLMEPM-v2TravelShop.CMBAvg.Hy-Memory-30B-A3B****73.2876.0053.2856.5619.0563.19****56.89Claude Opus 573.2880.4151.1945.9123.8949.5354.04DeepSeek V4.1 Flash72.5282.8349.7459.4718.8949.3855.47Hy368.7083.0048.6846.5219.4447.4452.30Qwen3.6-35B-A3B69.0879.0050.6043.5318.7743.5250.75Hy3-A3B Base64.5081.0052.1441.7420.5645.9350.98Qwen3-30B-A3B67.9460.0047.1044.8018.3341.2746.57

https://huggingface.co/tencent/Hy-Memory-30B-A3B#quickstartQuickstart

Deploy the model with vLLM (tested with v0.31.0) using\-\-served\-model\-name hy\-memory, then call the OpenAI-compatible API. The example below extracts user facts and preferences with no-think mode enabled.

from openai import OpenAI

client = OpenAI(
    base_url="http://127.0.0.1:8000/v1",
    api_key="EMPTY",
)

response = client.chat.completions.create(
    model="hy-memory",
    messages=[
        {
            "role": "system",
            "content": (
                "Extract explicit user facts and preferences from the conversation. "
                "Write each memory as a concise, self-contained statement. "
                "Do not infer or invent information. "
                'Return only a JSON object with a "memories" list of strings.'
            ),
        },
        {
            "role": "user",
            "content": (
                "My name is Alex. I live in Seattle and work as a software engineer. "
                "I mainly use Python and prefer concise answers with code examples. "
                "I'm vegetarian, so please keep that in mind when suggesting meals."
            ),
        },
    ],
    temperature=0.0,
    max_tokens=512,
    extra_body={
        "chat_template_kwargs": {
            "reasoning_effort": "no_think",
        },
    },
)

https://huggingface.co/tencent/Hy-Memory-30B-A3B#resource-availabilityResource Availability

• Framework Code & Evaluation Utilities:https://github.com/Tencent-Hunyuan/Hy-Memory
• Model Weights, Configuration, and Model Card:https://huggingface.co/tencent/Hy-Memory-30B-A3B
• Technical Report:https://arxiv.org/pdf/2610.22196