IROS 2026 best paper and BetaMem paper put long-term embodied memory at center of robot research

LT-Mem addresses what its authors call "temporal amnesia": systems either overwrite old states with new observations or store semantic snapshots without consistent object identities across visits. The paper uses a volatility score tied to each object's behavior to decide whether a new observation should overwrite, preserve or coexist with an earlier state. The authors acknowledged limits: tests covered two indoor scenes and one outdoor scene, with 10 visits each, and linking identities of visually similar objects remained difficult.

The report said the industrial need is broader. General-purpose robots must operate for hours, days or weeks, not minutes. Without long memory, every encounter is a first encounter and the robot remains an isolated executor of single actions; with it, the machine can become a long-term colleague. The technical problem is not just expanding a context window. Continuous multimodal data cannot all be stored or retrieved, the world changes, new observations cannot simply replace or pile onto old ones, and a false memory can be repeatedly recalled and amplify errors. How memory is written, evidenced and corrected are open questions, the report said.

It described four research strands. Spatial memory, represented by Hydra, ConceptGraphs, Khronos and DynaMem, records what is where; DynaMem updates memory after objects are moved and achieved an average 70% pick-and-place success rate on non-static objects in real-robot tests, more than twice that of static systems, according to the report. Detection memory, such as ReMEmbR, retrieves past observations by similarity but does not answer whether a cup seen five minutes ago is still on the table. Change memory, including LT-Mem, asks whether an object has changed. Operation memory, including MemoryVLA and MEM, tracks task progress; MemoryVLA reported a 26-percentage-point gain in real long-horizon success over its strongest baseline, while MEM used video for short-term memory and natural language for long-term memory and tracked tasks for up to 15 minutes. The report identified remaining gaps: spatial and operation memory rarely meet, memory management is often hand-designed, and most systems store descriptions rather than raw evidence that can be checked.

BetaMem, from Chinese startup Beita Wuxian, is presented as an attempt to close those gaps. The team posted its methods and experiments to arXiv on Oct. 20. According to the report, BetaMem uses the 9B open-source model Qwen3.5-9B to generate memories and claims state-of-the-art results on four benchmarks: M3-Bench-robot, LoCoMo, OpenEQA and EgoLifeQA. The report said its advantages appeared in tasks that require linking clues across observations, recalling events over long spans and identifying personal habits.

The architecture separates evidence from state. A temporal state graph records what the robot currently believes the world is like; an append-only event log records how that state changed; and a finite-capacity source archive keeps selected raw images and audio. The three share entity, event and source identifiers. Archive decisions, such as whether to keep a full frame, crop it, degrade it or delete it, are weighed against storage and read budgets. Each state record carries spatial context such as location or container, coordinate frame and map version, along with confidence, source, version number and a status of observed, inferred, conflicting, expired or unknown. It also records three times: the period for which the described fact is valid, when it was seen and when it was written. Uncertain transition times are stored as intervals, and anything beyond direct observation is marked as inference.

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 reference one another; an episodic state transition can revise an entity state and attach it to a place and map version through L1. L4 experience memory stores regularities, and L5 procedural memory holds task preconditions, callable skills, expected observations and recovery branches. The event log and source archive run across all six layers.

In the paper's key example, picking up a key, handing it over and putting it down are three separate events, each with an executor and links to timestamped images and audio. A speech transcript keeps the speaker and links back to the original audio. At retrieval, the robot can distinguish who took the key from who put it away, and can separate a stated intention from a reported result and from an observed change. Writing also has gates: observations are registered with source and collection time, cut into windows, and turned into entity, relation and event proposals with source citations. Proposals are checked against appearance, location, time continuity and existing identities; uncertain cases keep multiple hypotheses. Before a commit, the system checks source evidence, validity range and record versions. A commit appends an event and updates state, while the old state remains in history. The report said a historical record is treated as a search clue, not proof of the present: if the key was last seen in a drawer, the robot goes to the drawer, obtains a new observation and then appends a new state version.

The report said this is consistent with LT-Mem's conservative updating but broader, because BetaMem's state can include task progress, container open or closed status and action results as well as object positions. In May, co-founder Tao Shuai said in an interview that the lack of long-lasting, multimodal spatiotemporal memory was a priority. The team later released BetaMem, the world action model BetaWAM 0.1 and its first robot, Duomi. Tao has also said the memory system can reuse experience from his work on personalized recommendation systems, according to the report.

Within the team's unified brain, BetaBrain, the model, agent and action layers read the same state at different time scales. The action layer uses short-term action and perception memory: when an arm blocks the target or a robot turns a bottle cap several times, BetaMem compresses what has happened into a few updated variables, including the current phase, the last confirmed target pose and recent force feedback. Small deviations are absorbed at that layer; only deviations above a threshold are written back to task state to trigger a retry, a skill switch or an escalation. Cross-episode physical experience, such as lifting a cabinet door before pulling it or a wet glass being slippery, is stored as reusable object-level and environment-level entries for the action model and then distilled into policy. After an action, a world model predicts the outcome using the full remembered state, including parts not currently visible but previously confirmed, and a verifier compares the prediction with visual, force and pose feedback. Acceptable error commits a checkpoint; local deviation returns to the action loop; a failed hypothesis returns to BetaBrain for replanning. The threshold for abnormal error is learned from the object's and action's history rather than set by hand, the report said.

Beita Wuxian also places memory policy and behavior policy in one feedback framework and optimizes them jointly, with rewards covering task completion, key evidence coverage, state consistency, failure recovery, user satisfaction, execution efficiency, long-term gain and safety risk. To train memory formation, the team used supervised fine-tuning and GRPO reinforcement learning so a smaller model could associate current observations with history and record object states and changes. Rewards came from multiple independent records and raw-image review: evidence-supported clues that are easy to miss received higher weight, while fabrication, distortion and wrong associations were penalized. For policy improvement, the team designed failure-driven recursive self-improvement for memory policy. The system traces information retention, retrieval and timeliness judgments in a failure, proposes targeted rule fixes, keeps the base model and robot controller unchanged, and tests fixes through independent replay. Validated policies are used in later tasks, and new failures drive another round. The report described the process as versioned, rollback-capable and canary-tested, a rarity in embodied memory systems.

On M3-Bench-robot, BetaMem led the strongest reference by 15.96, 23.51 and 27.93 percentage points on multi-detail, multi-hop and character-understanding questions, respectively. An ablation that removed source evidence linked to original words dropped M3-Bench-robot accuracy from 48.08% to 38.61%, a loss of 9.47 points. In write learning, the share of generated write actions that conformed to memory-bank rules rose from 50.9% to 86.2%, above the teacher model's 72.0%, while output tokens per stored action fell from 224 to 118, a 47.3% reduction. The paper acknowledged shortcomings, the report said, while arguing that the direction, making memory trainable, traceable and rollback-capable, is what matters.