[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-mape-agent-memory-program-evolution":3,"topics-all":38,"news-related-1ec34a26-f012-4bc4-8193-14e651c07516":67},{"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},"1ec34a26-f012-4bc4-8193-14e651c07516","程序进化搜出记忆设计:涨12个百分点,成本只占35%","论文 MAPE 把智能体记忆系统当成可进化的程序来搜:从什么都不记的种子出发,在 LoCoMo 上搜出的方案准确率 87.94%,换成 Qwen 骨干后比手写系统高 12.01 个百分点、服务成本只占 35%;两条反复出现的规律是保留原始对话和按查询意图分流,后者移植进五个现成记忆系统后全部涨分。","智能体要「记住用户」，得有一套记忆系统：存下对话、检索相关片段、拼给模型回答。这两年它越做越复杂——抽取事实、知识图谱、离线整合、语义压缩。但两份独立工作先泼了冷水：一份论文报告，朴素 RAG 直接检索原始对话，在 LoCoMo、LongMemEval 等四个基准的七组设定里有六组排第一，领先被对照的人工记忆系统 0.13 到 5.53 个点（[Raw-RAG 论文](https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.30755)）；另一份则拆成「记忆收益」和「记忆税」，指出加工既可能带来收益，也可能抹掉有用证据（[Memory Tax 论文](https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.34073)）。\n\n那么记忆系统该怎么设计？一份 11 月 12 日提交 arXiv 的新论文给了条不一样的路：别让人设计，让程序自己搜。\n\n## 把手写设计换成程序进化\n\n论文叫《Discovering Effective Agent Memory Designs through Program Evolution》(https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.17068)，框架叫 MAPE。它把记忆系统拆成 build、retrieve、compose、answer 四个阶段，每个都可编程，在统一接口上做整程序进化：从一个「什么都不记」的最小种子出发，提议器改写整个程序，评测打分，诊断代理读执行轨迹找失败原因，再喂给下一轮提议。\n\n提议器和诊断代理都跑 Claude Opus 4.8，推理强度开到 xhigh。LoCoMo 上每个设置跑三轮，LongMemEval 上一轮，每轮 20 步。搜索集与评测集 20% \u002F 80% 切分，评测集在进化期间不可见。\n\n对照实验对象是五个手写记忆系统（Mem0、A-Mem、LightMem、SimpleMem、Zep）和朴素 RAG、Full Context。\n\n## 搜出来的是手写设计里没写出来的那部分\n\n进化结果有两套。LoCoMo 上搜出的 MAPE-ALCM 在 GPT-4.1 mini 上拿到 87.94%，同一张表里 Full Context 是 80.87%，SimpleMem 是 80.80%。LongMemEval 上搜出的 MAPE-ALME 用 Qwen3-30B-A3B-Instruct-2507 跑到 77.00%，该栏基线里朴素 RAG 是 66.00%。\n\n最抓眼的一组数字来自换成 Qwen 骨干：LoCoMo 上 MAPE-PLCM 拿到 81.33%，比 SimpleMem 的 69.32% 高 12.01 个百分点，服务成本 1.14 美元对 3.24 美元，只占 35%。\n\n可复用的不是分数，是两条反复出现的规律。\n\n第一条：**留着原始对话**。MAPE-ALME 抽取原子笔记的同时也索引原始对话轮次，检索时最多返回五轮原文。消融显示，去掉原始轮次，LoCoMo 从 75.90% 掉到 63.73%，LongMemEval 从 77.00% 掉到 70.50%；其中单会话助手类问题降幅达 20.00%。\n\n第二条：**按查询意图分流**。论文管它叫 Query-Routed Scaffolding，先按措辞把问题分到六种意图，再各自套专门的脚手架——时间问题做日期推算，聚合问题用代码去重计数。去掉它，LoCoMo 从 75.90% 掉到 71.46%，LongMemEval 从 77.00% 掉到 70.75%；两个都去掉，只剩 63.12% 和 63.00%。\n\n这两条规律与前面两份独立工作方向一致：加工会丢东西，原始证据不可替代。\n\n## 把规律塞进别人的系统里\n\n论文没停在自家分数：它把意图分流的组合模块移植进五个现成记忆系统，保留各自的构建与检索，只换证据怎么拼。\n\nLongMemEval 上五个系统全部涨分，涨幅在 2.75 到 9.25 个点：SimpleMem 从 49.25% 到 58.50%，Zep 从 46.75% 到 52.25%。\n\n幅度分化，论文给了原因：Mem0 只有 2.75%，它的失败主要不是「证据够了但没用好」，而是检索没把信息带回来——意图分流治的正是前者。错误归因分析印证：分流模块能修正 42% 到 60% 被判定为「证据充分但答错」的错误，对「证据就缺」的错误只有 3% 到 13%。\n\n## 把成本写进目标，计算就搬家了\n\n只用准确率当目标，搜索会一路往构建阶段堆东西——LongMemEval 轨迹上第 13 步引入「每会话抽一次笔记」的构建期操作，服务成本大约翻六倍。MAPE-ALME 最终约 92% 的服务成本花在构建期抽原子笔记上。\n\n把成本写进目标，搜出的程序形态完全不同：MAPE-PLME 直接用 BM25，构建期一次模型调用都不要；MAPE-PLCM 检索原始对话轮次，只在聚合类问题上才查抽出的事实。这两个程序的服务调用次数比纯准确率版本在 LongMemEval 上少 90%，在 LoCoMo 上少 40%。告诉优化器「省着花」，计算就从「入库时加工」搬到「查询时按需处理」。\n\n成本口径也写清楚：按固定单价折算成每千次查询的美元数，单价表日期是 2026 年 9 月 21 日，不含搜索代理调用与本地计算，不测延迟。\n\n## 换模型、换基准，排名会反转\n\n这套程序在迁移上很脆弱：同一 LoCoMo 上搜出的程序，服务模型换成 Qwen 骨干后排名直接对调——只追准确率的 MAPE-ALCM 在时间类问题上从 79.84% 掉到 43.80%，MAPE-PLCM 守住 70.93%。论文的解释是：前者把相对日期解析甩给问答模型，后者会在回答前把相对表述换成绝对日期。\n\n跨基准迁移也不总是赢。LoCoMo 上搜出的两个程序搬到 LongMemEval 是 65.50% 和 64.25%，压过五个手写记忆系统，却仍低于朴素 RAG 的 66.00%；反过来，LongMemEval 上搜出的程序搬到 LoCoMo 是 75.90% 和 74.67%，超过全部基线。\n\n## 论文自己承认的边界\n\n局限段写得很老实：结论只在两个问答基准上成立，组件分析只针对一个代表性程序；进化轨迹在不同 run 之间会变；成本结论还取决于构建成本能在多少查询间摊销。这是一份尚未标注会议出处的 arXiv 预印本。\n\n## 所以呢\n\n这份工作交付的不是一个更高的分数——分数会随基准和模型变——而是一套能自己找出设计的过程，加上两条被反复验证、能移植到别人系统的规律：留着原文，按意图分流。\n\n对做智能体的人来说，这比任何具体架构都值得抄：**构建阶段尽量少加工，加工留到查询时按问题类型来**。想省成本，就把成本写进搜索目标，别指望模型自己学会取舍。\n\n前提是评估口径要跟论文学：别只看准确率，还要看每次查询花了多少钱、证据齐不齐。","https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.17068","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},"120fa59a-ff6f-4537-9bf5-f818df636a0e","benchmark",{"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},"b455082d-513c-411a-840a-79b5245a4a76","en","Evolving Agent Memory: 12 Points Higher at 35% of Cost","Program evolution searches agent memory programs: 12 points over hand-built systems at 35% of serving cost, and intent routing lifts all five.","An agent that wants to remember its users needs a memory system: store the conversation, retrieve relevant snippets, compose them, and hand the result to the model. Over the past couple of years these systems have grown more elaborate, layering fact extraction, knowledge graphs, offline consolidation and semantic compression. Two independent lines of work recently pushed back. One paper reports that naive RAG, which retrieves directly over raw conversations, ranked first in six of seven benchmark-model settings across four benchmarks including LoCoMo and LongMemEval, beating the strongest agentic memory system it compared against by 0.13 to 5.53 points. Another splits the value of memory into memory benefit and memory tax, arguing that processing can remove or obscure useful evidence.\n\nA paper submitted to arXiv on 12 November 2026 takes a different route: rather than having humans design the memory system, let programs search for one.\n\n## Replacing hand-written design with program evolution\n\nThe paper is titled Discovering Effective Agent Memory Designs through Program Evolution, and the framework is called MAPE. It breaks a memory system into four programmable stages, build, retrieve, compose and answer, then runs whole-program evolution over a unified lifecycle interface. Starting from a minimal seed that stores nothing, a proposer rewrites the entire program, an evaluation scores it, and a separate diagnosis agent reads the execution traces to work out what went wrong, feeding that diagnosis into the next proposal.\n\nBoth the proposer and the diagnosis agent run Claude Opus 4.8 with reasoning effort set to xhigh. Each setting gets three evolutionary runs on LoCoMo and one on LongMemEval, with 20 steps per run and one candidate program proposed per step. Search and evaluation partitions are fixed at 20% and 80%, and the evaluation set stays completely unavailable throughout evolution.\n\nThe comparison set is five hand-designed memory systems, Mem0, A-Mem, LightMem, SimpleMem and Zep, alongside naive RAG and Full Context.\n\n## What the search found that hand-written designs did not\n\nThere are two resulting programs. On LoCoMo, the evolved MAPE-ALCM reaches 87.94% average accuracy with GPT-4.1 mini; in the same table Full Context sits at 80.87% and SimpleMem at 80.80%. On LongMemEval, the evolved MAPE-ALME reaches 77.00% with Qwen3-30B-A3B-Instruct-2507, where the strongest baseline in that column is naive RAG at 66.00%.\n\nThe most striking numbers come from swapping in the Qwen backbone. On LoCoMo, MAPE-PLCM reaches 81.33%, which is 12.01 percentage points above SimpleMem at 69.32%, while serving cost runs 1.14 dollars against 3.24, so cost accounts for only 35%.\n\nBut the reusable part is not the score, it is two design patterns that keep reappearing across the discovered systems.\n\nThe first is to keep the raw conversation. Alongside extracting atomic notes, MAPE-ALME also indexes the original conversation rounds and returns up to five of the highest-ranked raw rounds as evidence. Ablations show that removing the raw rounds drops LoCoMo from 75.90% to 63.73% and LongMemEval from 77.00% to 70.50%, with the largest single decline, 20.00%, on single-session assistant questions.\n\nThe second is to route by query intent. The paper calls this Query-Routed Scaffolding. Rules based on the query wording assign each question to one of six intents, temporal, aggregation, difference, current-value, preference or direct lookup, and each intent gets its own scaffold, such as date arithmetic for temporal questions and deduplicating counts in code for aggregation questions. Removing it drops LoCoMo from 75.90% to 71.46% and LongMemEval from 77.00% to 70.75%. Removing both leaves 63.12% and 63.00%.\n\nBoth patterns point the same direction as the two independent studies above: processing loses things, and raw evidence has value that cannot be replaced.\n\n## Transplanting the pattern into systems built by others\n\nThe paper does not stop at its own scores. It transplants the intent-routing composition module into five existing hand-designed memory systems, keeping the original construction and retrieval of each and changing only how evidence gets composed.\n\nOn LongMemEval all five improve, with gains between 2.75 and 9.25 points: SimpleMem goes from 49.25% to 58.50%, LightMem from 59.25% to 68.25%, Zep from 46.75% to 52.25%, and Mem0 from 47.50% to 50.25%.\n\nThe spread has an explanation. Mem0 gains only 2.75% because its failures are mostly not cases of evidence being available but unused, but rather retrieval failing to bring back the needed information in the first place, which is exactly what intent routing does not fix. The accompanying error attribution analysis confirms this: the routing module corrects 42% to 60% of errors judged recoverable, against only 3% to 13% of errors where the evidence was simply missing.\n\n## Put cost in the objective and computation moves\n\nA more immediately useful design choice is that the search objective can include serving cost alongside accuracy.\n\nWith accuracy as the only objective, search piles work into the build stage. On the LongMemEval trajectory, step 13 introduces per-session note extraction at build time and serving cost rises roughly sixfold. MAPE-ALME ends up spending about 92% of its serving cost on build-time atomic note extraction.\n\nOnce cost enters the objective, the discovered programs look completely different. MAPE-PLME uses BM25 and makes no model calls at build time at all, while MAPE-PLCM retrieves raw conversation turns and consults extracted facts only for aggregation questions. These two accuracy-cost programs make 90% fewer serving calls than their accuracy-only counterparts on LongMemEval and 40% fewer on LoCoMo.\n\nIn other words, telling the optimizer to spend carefully moves computation from processing at write time to selective processing at query time. The cost accounting is spelled out too: fixed unit prices converted into dollars per 1,000 queries, with a price list dated 21 September 2026, excluding judge calls, search-agent calls, local computation and storage, and with latency not measured.\n\n## Rankings flip when the model or benchmark changes\n\nThis is where the programs turn out to be fragile.\n\nServing the same LoCoMo-evolved programs with the Qwen backbone directly reverses the ranking between the two. Accuracy-only MAPE-ALCM falls from 79.84% to 43.80% on temporal questions, while MAPE-PLCM holds at 70.93%. The explanation given in the paper is that MAPE-ALCM leaves relative date resolution to the question-answering model, whereas MAPE-PLCM converts relative expressions such as last week into absolute dates before answering.\n\nCross-benchmark transfer does not always win either. The two programs evolved on LoCoMo reach 65.50% and 64.25% when moved to LongMemEval, beating the five hand-designed memory systems but still below naive RAG at 66.00%. In the other direction, programs evolved on LongMemEval reach 75.90% and 74.67% on LoCoMo, ahead of every baseline.\n\n## What the paper admits about its own limits\n\nThe limitations section is candid: the findings hold on two question-answering benchmarks, the component analysis centres on one representative program, evolutionary trajectories vary between runs, and the cost comparison covers serving token charges only, with conclusions that depend on how build-time cost amortizes across queries. This is an arXiv preprint with no venue noted, listing 24 pages, 8 figures and 10 tables.\n\n## So what\n\nWhat this work actually delivers is not a higher score, since 87.94% and 77.00% will both move with the benchmark and the model, but a process that can find designs on its own, plus two patterns that were validated repeatedly and can be transplanted into systems built by others: keep the raw text, and route by intent.\n\nFor anyone building agents, that is more worth copying than any particular architecture. Do as little processing as possible at build time, and move processing to query time, keyed by question type. If you want lower cost, put cost into the search objective rather than hoping the model learns to economise on its own.\n\nThe precondition is that evaluation has to follow the protocol used in the paper: not just final accuracy, but also what each query cost and whether the retrieved evidence was complete.","mape-agent-memory-program-evolution","2026-11-14T05:05:00Z","2026-11-14T05:16:37.822807Z","2026-11-14T05:16:37.822822Z",true,"agent",259,[39,49,58],{"slug":40,"tag_slug":41,"title_zh":42,"title_en":43,"intro_zh":44,"intro_en":45,"id":46,"is_active":35,"created_at":47,"modified_at":48},"image-generation","text-to-image","图像生成：从文生图到逐框编辑的模型竞赛","Image Generation: from text-to-image to frame-level editing","文生图已是红海，战场正转向编辑与可控性：FLUX.2\u002F3 用开源和量化把生图搬进消费级显卡，Qwen-Image、Seedream 等国产模型紧咬不放，GPT Image 2 则把推理能力带进像素。本专题追踪图像生成与编辑模型的发布、架构演进、加速量化与商业化数据，看「好图」如何变成「可用的图」。","Text-to-image is a crowded race, and the frontier has moved to editing and control: FLUX.2\u002F3 push open weights and quantization onto consumer GPUs, Qwen-Image and Seedream close in from China, and GPT Image 2 brings reasoning into pixels. This topic tracks model releases, architecture shifts, acceleration work, and the business numbers behind \"pretty\" becoming \"usable\".","52de7c59-7864-4569-8ae7-3d656d26a1a1","2026-11-20T07:20:23.534506Z","2026-11-20T07:20:23.534527Z",{"slug":50,"tag_slug":50,"title_zh":51,"title_en":52,"intro_zh":53,"intro_en":54,"id":55,"is_active":35,"created_at":56,"modified_at":57},"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":59,"tag_slug":59,"title_zh":60,"title_en":61,"intro_zh":62,"intro_en":63,"id":64,"is_active":35,"created_at":65,"modified_at":66},"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":68},[69,74,79,84,89,94],{"id":70,"title":71,"news_slug":72,"published_at":73},"417ea170-637f-48be-a998-aa13ec5cdb10","记忆税:11个记忆系统多数跑输朴素RAG","memory-tax-agent-memory-net-value","2026-10-28T21:07:37+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"00346b75-f071-42fd-ae16-db4c5569f01a","EarlyEval 提前叫停注定失败的 Agent:近半 token 省下,分辨率只动一两个点","earlyeval-early-stop-agent-eval","2026-09-03T21:04:52+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"10bd728a-5d05-461b-ba14-7c6ea51f8ce3","HyperTool 把 MCP 工具调用「折叠」成一段代码，Qwen3 智能体准确率直接翻倍","hypertool-mcp-folded-code-qwen3-2x","2026-06-14T20:00:00+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"c4623b85-37f5-435e-ae98-e67eb72eb1dc","别再逐篇问大模型:AI写程序当评委,每主题0.17美元起","execjudges-llm-written-program-judges-retrieval-eval","2026-11-20T01:40:00+00:00",{"id":90,"title":91,"news_slug":92,"published_at":93},"f02185e7-5f4c-46c6-bb18-575fbf3b1c0a","摩根大通实测:五项前沿配方,四项真有用","jpmorgan-factorial-retest-llm-ingredient-gains","2026-11-19T17:25:00+00:00",{"id":95,"title":96,"news_slug":97,"published_at":98},"03403319-8597-4876-8661-fad74ee0ac48","屏幕上的警告,智能体最高92%照做","computer-use-agents-cross-screen-harm","2026-11-18T23:30:00+00:00"]