Choose memory by the uncertainty it resolves.
A robot may need to remember an instruction, a hidden event, a previous action’s effect, or where a target moved. These facts have different lifetimes. Moving a cup invalidates its old position, while the instruction to wash it before placing it on a shelf can remain valid. The survey’s contextual evidence roles make this distinction explicit.
Episodic memory retains particular experiences. A semantic record summarizes what happened; a persistent skill makes an acquired procedure callable. Retrieval decides which stored evidence enters the current context. Its usefulness depends on both relevance and the executor’s ability to act on the result.
| Memory choice | Useful evidence | Decision requirement |
|---|---|---|
| Recent or recurrent state | Action effects, progress and observations within a sequence. | Compress history without losing a decisive event; reset when its physical assumptions change. |
| Episodic retrieval | Demonstrations, successful trajectories and failed attempts. | Match the task, physical conditions and execution phase, rather than appearance alone. |
| Event or scene records | Completed prerequisites, object locations and earlier interactions. | Update changed facts while retaining relevant evidence and its observation time. |
| Persistent skills and guidance | Callable routines, corrections and conditions of use. | Check arguments and preconditions; revise a routine separately from the advice selecting it. |
Four studies expose different memory interfaces.
- MemoryVLA combines perceptual and cognitive memory for manipulation. Historical evidence can resolve events absent from the current view. The survey uses it to distinguish retaining task state from inferring an unfamiliar task rule.
- ReCAP retrieves state–action chunks for a frozen predictor. Its main world-action-model branch jointly generates future observations and residual actions. Expanding the archive changes available evidence; the retrieved material still needs to support the current stage of execution.
- DROC distills language corrections into retrievable guidance for plans and skill calls. Here, distillation creates textual knowledge. A rule such as opening a drawer before reaching inside can be reused without training new neural weights.
- EmbodiedEvo separates parameterized Skills from condition-linked Lessons. Fresh test agents receive frozen artifacts acquired during practice. Comparing these artifacts with raw history examines how experience is represented for reuse, including its token cost.
These mechanisms span action inference, prediction and skill execution. Memory storage alone therefore does not determine a method’s place in the four-family taxonomy.
Test useful recall at a matched physical state.
Compare retained and cleared memory with the same current observation, task, executor and action budget. A hidden-cue task should require earlier evidence; a reactive control should remain solvable from the present view. Vary the delay between observing the cue and needing it, as well as distractors and storage limits.
Measure the full chain: recovery of the relevant record, a justified change in action, and physical task completion. Accurate retrieval can coexist with an executor that ignores the retrieved constraint. For cross-attempt reuse, reset the scene independently of memory and preserve the number of retries. Report repeated errors, corrections, retrieval latency and execution cost alongside success.
Make retained knowledge expire selectively.
A useful research direction is memory that records why evidence applies: object identity, contact conditions, task stage and executor version. Updating an object’s location should not erase its role in the taught procedure. Conversely, a successful grasp on one material may need rechecking after the surface changes.
Another opportunity is comparing read-time access to raw episodes with summaries formed when memories are written. A later question may require detail that an earlier summary discarded. Tests should vary the new information need while matching archive access, compute and interaction budgets.
Continue through the survey.
Read shared mechanisms: correspondence and memory, external knowledge reuse and revision, the memory mechanism comparison, and evaluation of improvement sources.
Related topics: robot self-improvement, navigation and spatial memory, and vision-language-action models.