[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-memorilla-latent-semantic-memory-tokens":3,"topics-all":38,"news-related-851c9f58-ab76-414d-86b4-4b10dc782422":57},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":24,"news_slug":31,"published_at":32,"created_at":33,"modified_at":34,"is_published":35,"publish_type":36,"image_url":14,"view_count":37},"851c9f58-ab76-414d-86b4-4b10dc782422","Memorilla:16个记忆token反超全上下文","斯坦福团队提出记忆系统 Memorilla：把整个文档库压成 16 个记忆 token，直接写进模型的输入嵌入空间，模型权重全程冻结，只训练记忆模块。论文报告事实验证基准 FactKG 准确率 84.33%，反超全上下文的 80.43%，文档 token 少用 108 倍。它把记忆从检索工程变成了训练工程。","把 512 篇文档原样塞进上下文，单卡 H100 上要跑 19.9 秒；换成 16 个“记忆 token”，只要 0.74 秒。更反直觉的是，论文报告在事实验证基准 FactKG 上，后者不但没掉分，反而从全上下文的 80.43% 涨到 84.33%。这套方法叫 Memorilla，论文已在 arXiv 公开（[论文](https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.21375)），代码与权重都以 Apache 2.0 开源。\n\n## 它不是压缩文本，是学出一套“软”记忆\n\n主流做法是检索：先挑几段原文塞进提示词。Memorilla 换了思路——一个冻结的编码器把每篇文档编码一次存进向量库；提问时，一批可学习的 memory slot 读入问题和整个文档库，直接产出与问题相关的记忆 token，写进大模型的输入嵌入空间里。\n\n关键点有两个。大模型权重全程冻结，只有记忆模块被训练；而且不管文档库多大，模型收到的 token 预算恒定。换句话说，输入侧成本与文档规模脱钩：文档再多，模型看到的始终是 16 个 token。团队把这种设计称为“把记忆当成一种模态”，作者的说法是“如果 RAG 是情景记忆，Memorilla 就是语义记忆”。\n\n## 数字：哪些赢，哪些输了\n\n在 8 个基准上，论文报告平均分 49.68，为所有方法中最高，并在 5 个基准上领先：\n\n- PersonaMem-v2 达到 44.60%，最强基线 xRAG 为 30.90%，全上下文只有 27.92%，而且多用 1974 倍文档 token\n- FactKG 84.33%，全上下文 80.43%，少用 108 倍 token\n- TriviaQA 精确匹配 57.00%，文档 token 少 743 倍\n- LaMP-7 的 ROUGE-L 为 48.37，LaMP-4 为 18.15\n\n也有输的地方，得一并说清：PubMedQA 上 59.00%，低于 LLMLingua-2 的 69.00%；PersonalizationV4 上 66.65%，略低于 MemoRAG 的 67.88%；NarrativeQA 上 19.31%，略低于 MemoRAG 的 20.51。团队自己的结论是：学出来的记忆赢在“需要把整块材料拼起来看”的题，检索赢在“答案是一个能原样查到的名字”，短文本上直接读原文仍然更好。\n\n延迟与压缩方面，单张 H100-80GB、batch size 1、Qwen3-8B，512 篇文档（约 12.8 万 token）下，Memorilla 稳定在 0.74 秒，而全上下文从 0.75 秒一路涨到 19.9 秒。NarrativeQA 的书籍和剧本被压缩超过 4000 倍，作者在社交平台给出的口径是最高 4212 倍 token 节省。八个基准平均下来，文档 token 少用 394 倍。\n\n## 智能体那一侧：记忆可以自己学出来\n\n同一套记忆模块被接到 TextWorld 文本冒险游戏里：一个冻结的 Qwen3-4B 智能体要在最多 65 个房间、50 步内完成任务，历史被压成 8 个记忆 token，只有记忆模块用 GRPO 强化学习训练。\n\n结果是：平均成功率从 18.6% 升到 24.2%，三次之内通关的比例从 25.0% 升到 37.5%，回合用尽前结束的比例从 29% 升到 42%。论文特别指出，未训练的随机记忆反而低于基线——增益来自学习，而不是来自“多塞了几个 token”。\n\n## 它补的是检索的空位\n\n更有意思的是两者并不重叠。FactKG 上，RAG-50 检索拿到 78.4%，Memorilla 是 84.3%，而两者取并集的上界高达 94.3%：15.9% 的题只有 Memorilla 能解，10.0% 只有检索能解。论文强调这个 94.3% 是 oracle 上界、并非实测的混合系统，但它指明了方向——记忆负责读懂材料之间的关系，检索负责把名字原样找出来。Hacker News 上一位读者把机制概括成一句话：注入进去的更像一种“vibe”而不是可引用的原文，这恰好对上了它在概念题上强、在逐字检索题上弱的表现。\n\n## 现在的边界\n\n三条限制写在项目主页上，值得原样记住。第一，它会丢细节，比如确切的名字。第二，因为记忆直接写进输入嵌入，方案要求开放权重，而且每个 LLM 都要单独训一个记忆模块。第三，全部实验都是 Qwen 编码器配 Qwen 解码器、只用文本，跨模型家族与跨模态都还没验证，TextWorld 也只测了单一游戏套件、最多 50 步。\n\n训练成本也不算轻：三个阶段在 4 张 H100 80GB 上跑约 19 小时，要下载约 310 GiB 数据与嵌入，评测全部八个基准在单卡上约 15 分钟。代码仓库截至抓稿时有 339 个星、26 个 fork，权重则以 Apache 2.0 发布在 Hugging Face 上。\n\n## 所以呢\n\nMemorilla 最值得注意的地方，可能不是那 16 个 token，而是它把记忆从检索工程变成了训练工程：你要的不是“更好的召回”，而是一套能学会怎么记的参数。门槛也很直白——需要开放权重，需要为每个模型付一次训练钱，目前只在文本上验证过。\n\n对正在做长上下文与 Agent 记忆的团队，值得抄的也许不是整套系统，而是那个固定 token 预算：让模型无论面对 512 篇还是 50 万篇文档，输入侧成本都不再随规模增长。接下来真正的问题是，当记忆可以被训练，谁来训练它、又是为谁训练。\n","https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.21375","7437aeb9-930c-4866-a2e9-48003c1a792b",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"6ad31a14-c0da-42df-81fd-564281f768db","agentic-ai",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"2d9c2fb0-2be5-4ad1-aedb-e9747addf355","compression",{"id":19,"name":20,"slug":20,"description":14,"color":14},"0ef8513a-0a26-42f0-b6f9-5b6dadded45c","efficiency",{"id":22,"name":23,"slug":23,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"19657ac2-6fce-4f46-9af4-7c487b4926a0","en","Memorilla: 16 memory tokens beat full context","Memorilla compresses a document collection into 16 memory tokens for a frozen LLM, reporting 84.33% on FactKG versus 80.43% for full context.","Pushing 512 documents into the context window takes 19.9 seconds on a single H100. Handing the model 16 memory tokens instead takes 0.74 seconds. And on the fact-verification benchmark FactKG, the compressed version does not merely hold its ground: it climbs from 80.43% with full context to 84.33%. The method is called Memorilla. The paper is public on arXiv (paper: https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.21375), and both the code and the weights are released under Apache 2.0.\n\n## Not compression, a learned soft memory\n\nThe mainstream approach is retrieval: pick a few passages and paste them into the prompt. Memorilla inverts that. A frozen encoder encodes each document once and stores it in an embedding store. At query time, a set of learned memory slots reads both the question and the entire collection, then emits query-dependent memory tokens directly into the input embedding space of the language model.\n\nTwo properties matter. The LLM weights stay frozen throughout, and only the memory module is trained. And the token budget handed to the model is fixed, no matter how large the collection grows. Put differently, input-side cost decouples from document count: however many documents are stored, the model still sees 16 tokens. The team describes the design as treating memory as a modality. One of the authors frames it as: if RAG is episodic memory, Memorilla is semantic memory.\n\n## The numbers: where it wins, where it loses\n\nAcross eight benchmarks, the paper reports a highest mean score of 49.68 and leads on five of them:\n\n- PersonaMem-v2 reaches 44.60%, against 30.90% for the strongest baseline xRAG; full context manages only 27.92%, and it does so while using 1974 times more document tokens.\n- FactKG reaches 84.33%, versus 80.43% for full context, using 108 times fewer tokens.\n- TriviaQA exact match reaches 57.00% with 743 times fewer document tokens.\n- LaMP-7 scores 48.37 ROUGE-L and LaMP-4 scores 18.15.\n\nThere are losses too, and they deserve to be stated plainly: PubMedQA comes in at 59.00%, below LLMLingua-2 at 69.00%; PersonalizationV4 lands at 66.65%, slightly under MemoRAG at 67.88%; NarrativeQA reaches 19.31%, slightly under MemoRAG at 20.51. The team's own reading is that learned memory wins on questions that require fitting a whole body of material together, while retrieval wins when the answer is a name that can be found verbatim, and that on short sources simply reading the text directly is still better.\n\nOn latency and compression, with one H100-80GB at batch size 1 running Qwen3-8B, 512 documents (about 128K tokens) leave Memorilla steady at 0.74 seconds, while full context climbs from 0.75 seconds all the way to 19.9 seconds. Books and movie scripts from NarrativeQA are compressed by more than 4,000x, and the author gives a personal ceiling of 4,212x fewer tokens on social media. Averaged across the eight benchmarks, the method uses 394 times fewer document tokens.\n\n## The agent side: memory that learns itself\n\nThe same memory module was hooked into the TextWorld text-adventure environment. A frozen Qwen3-4B agent has to complete a quest across up to 65 rooms within 50 turns, its history compressed into 8 memory tokens, and only the memory module is trained, using GRPO reinforcement learning on the reward the game provides.\n\nThe results: average success rises from 18.6% to 24.2%, the share of games finished within three attempts rises from 25.0% to 37.5%, and the share of episodes ending before the turn limit rises from 29% to 42%. The paper makes a point of one control condition: untrained random memory falls below the baseline. The gain comes from learning, not from having a few more tokens in the prompt.\n\n## It fills the gap retrieval leaves\n\nMore interesting is that the two approaches do not overlap. On FactKG, RAG-50 retrieval reaches 78.4% and Memorilla reaches 84.3%, while an oracle union of the two reaches 94.3%: 15.9% of questions are solved by Memorilla alone, and 10.0% only by retrieval. The paper stresses that the 94.3% is an oracle upper bound rather than a measured hybrid, but it points the way. Memory handles reading the relationships between materials; retrieval handles finding a name exactly as written. One Hacker News reader summarized the mechanism as injecting something more like a vibe than quotable text, which lines up neatly with a method that is strong on concepts and weaker on exact-text retrieval.\n\n## Where the boundary sits now\n\nThree limitations are written on the project page and are worth repeating. First, it loses fine-grained details such as exact names. Second, because memory is written directly into the input embeddings, the approach requires open model weights, and a memory module has to be trained for each LLM. Third, every experiment pairs a Qwen encoder with a Qwen decoder on text alone, so transfer to other model families and to other modalities is untested, and TextWorld covers a single game suite with episodes of at most 50 turns.\n\nTraining is not cheap either. The three stages take about 19 hours on 4 H100 80GB GPUs, downloading roughly 310 GiB of data and embeddings, and evaluating a checkpoint on all eight benchmarks takes about 15 minutes on one GPU. At the time of writing the repository had 339 stars and 26 forks, and the weights are published on Hugging Face under Apache 2.0.\n\n## So what\n\nThe most interesting part of Memorilla may not be those 16 tokens but what it does to the framing of memory. It moves memory from retrieval engineering to training engineering: what you want is no longer better recall, but a set of parameters that learn how to remember. The price is equally direct. You need open weights, you have to pay for training a module per model, and so far everything has been validated on text alone.\n\nFor teams working on long context and agent memory, the part most worth copying may not be the whole system but the fixed token budget: letting the model see the same small budget whether the collection holds 512 documents or half a million. The real question that follows is simpler. When memory becomes trainable, who trains it, and for whom.\n","memorilla-latent-semantic-memory-tokens","2026-11-03T07:03:31Z","2026-11-07T09:15:26.512976Z","2026-11-07T09:15:26.512984Z",true,"agent",58,[39,48],{"slug":40,"tag_slug":40,"title_zh":41,"title_en":42,"intro_zh":43,"intro_en":44,"id":45,"is_active":35,"created_at":46,"modified_at":47},"ai-for-science","AI for Science 2026：从 UniPert 到 GPT-Rosalind 的硬核进化","AI for Science 2026: from UniPert to GPT-Rosalind","生命科学、化学材料、物理世界模型——AI 正在从\"语言工具\"变成\"实验伙伴\"。本专题收录 AI 在三大科学方向的关键节点：UniPert 统一基因与化学扰动空间、GPT-Rosalind 端到端生命科学推理、达摩院 AI 智能体 28 小时找到 4 种超导新材料、Anthropic Claude Science 把工作台做成标准品。","From language tool to lab partner — AI is reshaping life sciences, chemistry\u002Fmaterials, and physical world models. This topic covers the key milestones: UniPert unifying genetic-chemical perturbation spaces, GPT-Rosalind's end-to-end life-sciences reasoning, DAMO's AI agent discovering 4 superconducting materials in 28 hours, and Anthropic's Claude Science workbench going mainstream.","988a4300-5fab-41c4-b5d8-63711a2dc757","2026-09-10T01:34:15.296649Z","2026-09-10T01:34:15.296663Z",{"slug":49,"tag_slug":49,"title_zh":50,"title_en":51,"intro_zh":52,"intro_en":53,"id":54,"is_active":35,"created_at":55,"modified_at":56},"h3-series","MiniMax H3 系列：从开源权重到 35 倍吞吐","MiniMax H3 Series: from open weights to 35x throughput","MiniMax H3 自 2026 年 8 月开源以来节奏密集：官方把生成、参考与编辑收回一个模型；ComfyUI 当天压进 RTX 3060；摩尔线程 3 小时完成国产 GPU 适配；fal 后训练版把吞吐拉到 35 倍；FastH3 蒸馏再砍推理成本。本专题持续追踪 H3 的发布—开源—蒸馏—部署全链路。","Since MiniMax open-sourced H3 in August 2026 the pace has been relentless: one unified omni-modal model, same-day ComfyUI support down to an RTX 3060, a 3-hour Day-0 port to Moore Threads GPUs, fal's post-trained H3 Max at 35x throughput, and FastH3 distillation cutting inference cost further. This topic tracks the full H3 chain — release, open weights, distillation, deployment.","83ef0daa-3c31-4cb3-86ed-e5ee58654d5f","2026-09-08T07:33:19.942193Z","2026-09-08T07:33:19.942209Z",{"items":58},[59,64,69,74,79,84],{"id":60,"title":61,"news_slug":62,"published_at":63},"f28966de-08b7-442e-ba3a-1d644131df26","工具输出不必是人话:清华潜token压缩最多省89%","coca-cola-latent-tool-context-compression","2026-11-07T23:05:00+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"821cbcc4-a8e0-41fe-90a8-284abd96f641","Agent上下文压缩提速8.68倍:小模型起草大模型验收","speccompact-agent-context-compaction","2026-11-05T19:20:00+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"a820bb5a-ac48-4aab-8d8b-34da340900c1","REMORY 让 LLM 用 5% 上下文继续跑 Agent","remory-agent-memory","2026-10-11T09:00:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"b8ca86ba-2401-441d-ba5c-06cb6ef09420","随机抽样几乎就够了:高误差风险再砍50.2%","bestrand-benchmark-compression-random-sampling","2026-11-06T15:40:00+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"53d00d00-e4fb-41b0-a36a-37161203b514","KV缓存拼接复用丢题可至100%,修补要盯答案块","kv-cache-repair-budget-target-chunk","2026-11-05T15:09:02+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"64088ca6-c244-474c-93ce-cdbfe55de6d6","投机解码没用满?agent循环把加速比挖到10.87倍","specpilot-agentic-speculative-decoding","2026-11-04T13:15:17+00:00"]