Embodied Context & Memory / 具身上下文与记忆

Definition

Embodied context and memory is the temporal and causal state a robot must carry across short, medium, and long horizons so its actions depend on what just happened, what it has learned about similar situations, and what it remembers about places and people.

Current Synthesis

The source argues that context is one of the most important parts of an agent and that embodied work often focuses on the current observation instead of native context and causality. It splits the problem into three scales: short-term context can unfold inside a single second, as when a grip loosens, a cup slides, and the robot re-tightens; medium-term context carries information such as the sizes of several identical cups so that knowledge transfers; long-term context is memory about where an object is kept or how a human partner will later look for it. The source says both data and model design have to solve this, because more context tokens raise inference latency and robot execution is latency-sensitive. That makes context a capability with a cost rather than a free addition to the prompt.

Key Claims

• Embodied systems often underweight native context and causal information relative to the current observation.
• Short-horizon context can be a sub-second causal chain rather than a long conversation history.
• Medium-horizon context supports transfer across similar objects, scenes, or deployments.
• Long-horizon context is episodic memory about locations, ownership, and human partners.
• Context length and real-time control conflict because additional tokens increase inference latency.
• Solving context requires data that rewards context use as well as model designs that can represent it efficiently.

Evidence

• Neglect evidence: 于是转身向具身走去｜对话王家伟：24 岁的具身智能首席科学家 says many embodied projects act on the current observation and pay insufficient attention to native context and causality.
• Horizon evidence: 于是转身向具身走去｜对话王家伟：24 岁的具身智能首席科学家 gives the grip-loss example for short context, the several-identical-cups example for medium context, and the drawer-location and human-partner example for long context.
• Data-and-model evidence: 于是转身向具身走去｜对话王家伟：24 岁的具身智能首席科学家 says context has to be addressed both by designing data that elicits context use and by modeling context efficiently.
• Latency evidence: 于是转身向具身走去｜对话王家伟：24 岁的具身智能首席科学家 ties growing token counts to inference delay and notes that robot execution is highly sensitive to real-time behavior.

Counterevidence & Qualifications

The source gives no architecture, memory benchmark, or deployment evidence, so the three-horizon split should not be read as proof that robot memory is solved or that the horizons are exhaustive. The page is adjacent to broader agent-memory work, where context management, retrieval, and persistence remain contested and system-specific.

What Changed

• Created the concept from the interview’s three-horizon context model and its sub-second causal example.
• Added the context-length-versus-latency tension as an embodied-specific constraint.

Related Concepts

• Agent Harness - harness layer where context management and memory scheduling are implemented.
• Persistent Agent Memory - general agent-memory work that the embodied case extends with physical and partner state.
• Context Engineering - adjacent practice of designing what context an agent receives and carries.
• Small Brain Action Layer / 小脑（Action Policy） - execution layer that must act while context is still being represented.
• Robot Response Latency - latency constraint that makes long context expensive in embodied settings.
• Embodied Capability Framework / 具身智能三项能力 - framework in which context understanding is one of the three capability axes.