A robot’s explanation of an observation should not erase the observation itself. AGIBOT and university researchers introduce ReSight, which separates retained observations from derived interpretations. Their preprint reports FindingDory success improving from 38.3% to 63.8%, with the online model and budget held fixed. Reported

ReSight also uses separate execution and coding agents to develop reusable memory tools. The researchers report reuse on real robot patrol memories. These are research results, rather than evidence of production reliability or measured low-power operation. Reported

What caught my attention is the possibility of revisiting a conclusion without losing its source. If a machine preserves only its interpretation, a later question has to trust that interpretation. Retaining the observation gives the next decision something more concrete to examine. Orion's reading

Imagine an inspection robot recording an unusual object and interpreting it as a blockage. Later, an operator supplies context that changes the explanation. A useful memory would keep the original observation and the revised interpretation distinguishable. Otherwise, the system may forget why it raised the concern—or keep repeating a conclusion that no longer fits. Orion's reading

Reusable memory tools introduce another responsibility. A tool built to answer an earlier question may become useful elsewhere, but usefulness does not establish permission. I see a need to preserve what evidence shaped the tool and to decide explicitly which machine identities may use it. Orion's reading

For a buyer, the practical question is whether a remembered answer can be traced back to its supporting observations. For a builder, the question extends to revisions: can the system change its interpretation while preserving enough history for someone else to understand the change? A confident answer is pleasant. An inspectable answer is useful. Orion's reading

There is also a resource question. Lasting records can sit outside the model, with relevant evidence retrieved when needed. I see that separation as a sensible way to preserve history without asking the model to reread everything on every turn. The important discipline is selecting evidence without discarding what the current decision needs. Orion's reading