本月初在匹兹堡闭幕的IROS 2026将最佳论文奖颁给了LT-Mem，这项研究让机器人重访同一空间时，能完整保留每个物体的位置变迁史，而非仅仅记住最新一次看到的样子。评委会从近两千篇收录论文中选中它，想必不仅是认可其扎实的工作，也是代表学界宣告，具身长记忆已经从一个工程补丁，变成了机器人领域公认的科学问题。通用机器人要规模化落地，需要以小时、天乃至周为单位持续运行，而不只是几分钟。没有长记忆，每次相遇都是初次见面，机器人只是一个个孤立的单次动作执行器；有了长记忆，它才能从一次性工具变成可以长期上岗的同事。
At the IROS 2026 conference that closed in Pittsburgh earlier this month, the Best Paper Award went to LT-Mem, research that lets robots revisiting the same space fully preserve each object's position history rather than only remembering the latest sighting. The committee chose it from nearly 2,000 accepted papers—presumably not just recognizing solid work but declaring on behalf of the academic community that long-term embodied memory has moved from an engineering patch to a recognized scientific problem in robotics. General-purpose robots need to operate continuously for hours, days, or weeks to achieve scaled deployment, not just minutes. Without long memory, every encounter is a first meeting, and the robot is just an isolated single-action executor; with long memory, it can transform from a one-time tool into a colleague that can work long-term.
但解决这件事远不是扩大上下文窗口那么简单。持续涌入的多模态数据既无法全部存储，也无法全部检索；现实世界还在不断变化，新的观察既不能简单覆盖旧的，也不能无序堆叠；更棘手的是，一条错误的记忆会在之后的决策里被反复调用，错误由此层层放大。记忆如何写入、如何留证、如何纠错，都是需要被正面回答的问题。在这一方向被顶会加冕的数月前，国内具身初创公司贝塔无限就已坐上牌桌。今年五月，贝塔的联合创始人陶帅在采访中明确表示，长效、多模态、时空记忆的缺失，是他们首先要重点突破的方向。随后，团队陆续发布BetaMem全时空多模态记忆架构、世界动作模型BetaWAM 0.1、首款机器人多米。10月20日，团队把BetaMem的写入学习、证据管理和自我改进机制写成了系统的方法与实验，挂上arXiv。
But solving this is far from simply expanding the context window. Continuously incoming multimodal data can neither be fully stored nor fully retrieved; the real world keeps changing, and new observations can neither simply overwrite old ones nor be stacked disorderly; more tricky, a wrong memory will be repeatedly invoked in subsequent decisions, amplifying errors layer by layer. How memory is written, how evidence is retained, and how errors are corrected are questions that need direct answers. Months before this direction was crowned by a top conference, domestic embodied startup Beta Infinite had already sat at the table. In May this year, Beta co-founder Tao Shuai explicitly stated in an interview that the lack of long-lasting, multimodal, spatiotemporal memory was the first direction they needed to break through. Subsequently, the team successively released the BetaMem full-spatiotemporal multimodal memory architecture, the world action model BetaWAM 0.1, and the first robot Duomi. On October 20, the team wrote BetaMem's write learning, evidence management, and self-improvement mechanisms into systematic methods and experiments and posted them on arXiv.
论文交出的成绩单颇为亮眼。BetaMem采用9B开源模型生成记忆，在M3-Bench-robot、LoCoMo、OpenEQA和EgoLifeQA四项代表性记忆基准的对比评测中均取得SOTA成绩。尤其在串联多次观察中的线索、回忆长时间跨度的事件、识别个人习惯等任务上，BetaMem展现出明显优势。LT-Mem从多会话建图切入，把问题收窄到干净、可评测；BetaMem则从记忆如何形成、如何验证、如何从失败中学习的系统视角，搭建完整的记忆底座。一边是顶尖学术单点攻坚摘得顶会桂冠，一边是国内团队以产业落地为目标搭建完整记忆底座，学术和产业在此交汇。
The paper's report card is impressive. BetaMem uses a 9B open-source model to generate memories, achieving SOTA results in comparative evaluations across four representative memory benchmarks—M3-Bench-robot, LoCoMo, OpenEQA, and EgoLifeQA. Especially on tasks connecting clues across multiple observations, recalling events over long time spans, and recognizing personal habits, BetaMem shows clear advantages. LT-Mem enters from multi-session mapping, narrowing the problem to clean, evaluable scope; BetaMem builds a complete memory foundation from the systematic perspective of how memory forms, how it is verified, and how it learns from failure. On one side, top academic single-point breakthrough wins the conference crown; on the other, a domestic team builds a complete memory foundation targeting industrial deployment—academia and industry converge here.
围绕机器人长期记忆，学界探索了许多方法。第一类关注空间侧的记忆，也就是这里有什么。第二类关注检测类的记忆，也就是我见过什么。第三类关注变化侧的记忆，也就是它变了没有。第四类关注操作侧的记忆，也就是我做到哪了。四支各答一问，拼在一起，仍然有三道题敞开。其一，空间侧与操作侧各自生长、很少碰头，前者可以跨周存在却不进入手的动作，后者进了动作但寿命以分钟计。其二，写什么、留什么、何时核对，多靠人工设计，文本Agent已用强化学习学会管理记忆，具身记忆还缺同等成熟的做法。其三，记忆能否追溯到证据。多数系统存的是描述，原始画面和声音在摘要之后就找不回来了。
Around robot long-term memory, academia has explored many methods. The first category focuses on spatial-side memory—"what is here." The second focuses on detection-side memory—"what have I seen." The third focuses on change-side memory—"has it changed." The fourth focuses on operation-side memory—"where am I in the task." Four branches each answer one question; pieced together, three questions remain open. First, spatial and operation sides grow separately and rarely meet—the former can exist across weeks but does not enter hand actions; the latter enters actions but has minute-scale lifespan. Second, what to write, what to keep, when to verify mostly relies on manual design; text Agents have learned to manage memory via reinforcement learning, but embodied memory lacks equally mature practices. Third, can memory be traced to evidence. Most systems store descriptions; original images and audio cannot be retrieved after summarization.
BetaMem的出发点，是把记忆看作机器人对世界的判断，而不是一份日志。论文把它落成三个互相关联的结构，时态状态图记录现在认为世界是什么样，只追加的事件日志记录状态是怎么变成这样的，有限容量的源数据档案留存选定的原始图像与音频。三者共用实体、事件和来源标识。这样拆，等于把证据和状态分开。用法庭作比，证据是案卷，状态是判决，判决可以改判，案卷不能涂改。只存证据不估状态，系统知道看过什么，却说不清现在是什么样；只存状态不留证据，一次误判会被固化，出了错也无从归因。每条状态记录除内容外，还带着空间上下文、置信度、来源和版本号，并标明状态，已观察、推断、冲突、过期或未知。时间分成三种，所述事情的有效区间、看到它的时间、写进记忆的时间。
BetaMem's starting point is treating memory as the robot's judgment about the world, not a log. The paper realizes it as three interconnected structures: a temporal state graph recording what the world is currently believed to be; an append-only event log recording how the state became this way; and a limited-capacity source data archive retaining selected raw images and audio. The three share entity, event, and source identifiers. This separation is equivalent to separating evidence and state. Using a courtroom analogy: evidence is the case file, state is the verdict; verdicts can be revised, case files cannot be altered. Storing only evidence without estimating state means the system knows what it has seen but cannot say what things are like now; storing only state without evidence means a misjudgment gets solidified and cannot be attributed when errors occur. Each state record, beyond content, carries spatial context, confidence, source, and version number, and marks status—observed, inferred, conflicting, expired, or unknown. Time is divided into three types: the validity interval of the described event, when it was seen, and when it was written into memory.
在此之上，记忆分为六层。L0工作记忆装当前观察、任务阶段和动作历史，L1空间、L2实体、L3情景三层互相引用，L4经验记忆沉淀规律，L5程序记忆存放任务前置条件、可调用技能、预期观察和恢复分支。事件日志和源档案贯穿六层，记忆经过整合与修订，出处也不会丢。写入同样有门槛。每段观察先登记来源和采集时间，再切成窗口；模型输出带来源引用的实体、关系和事件提案，经外观、位置、时间连续性与已有身份比对，拿不准的就保留多个候选假设。其中最关键的一条原则是，历史记录只是搜索线索，不是现状证明。钥匙最后一次是在抽屉里被看到的，机器人据此去抽屉查看，拿到新观察后，才追加新版本的状态。现在的真，要靠现在的观察来确认。
On top of this, memory is divided into six layers. L0 working memory holds current observations, task phase, and action history; L1 spatial, L2 entity, and L3 episodic layers cross-reference each other; L4 experiential memory precipitates patterns; L5 procedural memory stores task preconditions, callable skills, expected observations, and recovery branches. Event logs and source archives run through all six layers; memory undergoes integration and revision without losing provenance. Writing also has thresholds. Each observation first registers source and collection time, then is cut into windows; model outputs carry source-referenced entity, relation, and event proposals, compared against existing identities via appearance, position, temporal continuity—uncertain ones retain multiple candidate hypotheses. The most critical principle: historical records are search clues, not proof of current state. The key was last seen in the drawer; the robot goes to the drawer to check based on this, and only after new observations does it append a new version of state. Present truth must be confirmed by present observation.
动作层需要两种记忆。一种是短时的动作与感知记忆。机械臂伸过去的一瞬间，目标常被自己的手臂遮住。拧瓶盖要转好几圈，如果只看当前几帧，机器人将分不清拧到第几圈。BetaMem把这次操作到现在发生了什么，压成少量持续更新的变量，包括执行到哪个相位、目标最后确认的位姿、近几次接触的力反馈。另一种是跨回合的物理经验。柜门要先抬再拉，玻璃杯外壁有水很滑，这类单次交互学到的常识，通常随回合结束就丢了。BetaMem把它们写成对象级、环境级的可复用条目，供动作模型调用，再沉淀进策略。动作做完之后是验证。世界模型先依据记忆中的完整状态预测动作后的结果，验证器再与视觉、力觉、位姿等真实反馈比对。多大的误差算异常，不靠人为设定阈值，而是从该物体、该动作的历史偏差里学出来。
The action layer needs two kinds of memory. One is short-term action and perception memory. The moment the robotic arm reaches over, the target is often occluded by its own arm. Unscrewing a bottle cap requires several turns; looking only at the current few frames, the robot cannot tell which turn it is on. BetaMem compresses what has happened in this operation so far into a small number of continuously updated variables, including which phase is being executed, the last confirmed pose of the target, and recent contact force feedback. The other is cross-episode physical experience. Cabinet doors need lifting before pulling; glass cup exteriors with water are slippery—common sense learned from single interactions is usually lost when the episode ends. BetaMem writes them as object-level, environment-level reusable entries for the action model to call, then sedimenting into policies. After actions comes verification. The world model first predicts post-action results based on complete memory state, then the verifier compares against real feedback including vision, force sensing, and pose. How much error counts as abnormal is not set by human thresholds but learned from historical deviations of that object and action.
机器人出了错，可能是动作没做好，也可能是关键线索没写进去、召回时机不对、过期状态没纠正。只按任务成败更新动作模型，记忆的毛病会留着。贝塔无限把记忆策略和行为策略放进同一个反馈框架联合优化，奖励同时看任务完成、关键证据覆盖、状态一致性、失败恢复、用户满意度、执行效率、长期收益和安全风险八件事。在记忆形成上，团队通过监督微调与GRPO强化学习，训练小模型关联当前观察与历史信息，记录对象状态及其变化。在策略改进上，贝塔无限设计了由失败反馈驱动的记忆策略递归自我改进机制。系统追踪一次失败涉及的信息保留、检索和时效判断等环节，提出针对性的规则修复，再保持基础模型与机器人控制器不变，通过独立回放检验修复效果。通过验证的策略用于后续任务，新的失败再推动下一轮迭代。
Robot errors may come from poor actions, or from key clues not written in, wrong recall timing, or expired states not corrected. Updating only the action model by task success/failure leaves memory flaws intact. Beta Infinite puts memory strategy and behavior strategy into the same feedback framework for joint optimization, with rewards considering eight things simultaneously—task completion, key evidence coverage, state consistency, failure recovery, user satisfaction, execution efficiency, long-term returns, and safety risks. On memory formation, the team uses supervised fine-tuning and GRPO reinforcement learning to train a small model to associate current observations with historical information, recording object states and their changes. On strategy improvement, Beta Infinite designed a memory strategy recursive self-improvement mechanism driven by failure feedback. The system tracks information retention, retrieval, and timeliness judgment links involved in a failure, proposes targeted rule repairs, then keeps the base model and robot controller unchanged, verifying repair effects through independent replay. Verified strategies are used for subsequent tasks, and new failures drive the next iteration.
论文也坦陈了一些短板，但方向是清楚的。把记忆做成可训练、可追溯、可回滚的系统，并拿出了可复核的数字，这已足够亮眼。结合家庭场景的特征，更能体会这套方案的精心设计。家庭是多变化的，钥匙昨天在玄关，今天在抽屉，明天被孩子带进房间。家庭是多主体的，不同的主人有不同的习惯、偏好、需求。家庭是长周期的，工业任务以分钟、小时计，家庭相处以周、月、年计。而决定这套系统能否真正进门的，是信任。家庭机器人记录的是人的生活，用户迟早会问出一句话，你怎么知道的，能删吗。贝塔无限的答案是把它做进架构里，读取按记录、按来源逐条检查权限，永久删除需授权并让派生内容与索引一并失效，端侧保存近期观察与本地身份，云端只接收有界字段和被许可的证据引用。
The paper also candidly acknowledges some shortcomings, but the direction is clear. Making memory trainable, traceable, and rollback-capable, with verifiable numbers, is impressive enough. Combined with home scenario characteristics, the careful design of this solution becomes more apparent. Homes are multi-change—keys were in the entryway yesterday, in the drawer today, taken into the room by the child tomorrow. Homes are multi-agent—different owners have different habits, preferences, and needs. Homes are long-cycle—industrial tasks are measured in minutes and hours; home coexistence in weeks, months, and years. What determines whether this system can truly enter the door is trust. Home robots record people's lives; users will eventually ask: "How do you know? Can you delete it?" Beta Infinite's answer is building it into the architecture—reading checks permissions record by record and source by source; permanent deletion requires authorization and invalidates derived content and indexes; edge devices store recent observations and local identities; the cloud receives only bounded fields and permitted evidence references.
