what the model can see

part of the agent runtime lab.

an archive can hold every page you have ever written. the desk still holds the pages you chose to put there.

an agent runtime makes a similar choice each time it prepares a model request. somewhere there is a record of what happened. somewhere else there is the input being sent next. something has to decide how one becomes the other.

i built a small local runtime to make that choice visible.

first, i gave it a codeword:

1remember codeword=amber

the runtime saved the turn. then i asked for the codeword with zero earlier turns selected for the next request. the record still contained amber, but the request did not. the result said the codeword was missing from the supplied context.

select the earlier turns again, and amber returns.

this demonstration uses a deterministic offline fixture. its recall behavior comes from a rule that examines the supplied input. no hosted model was called, and this is not a measurement of a model's ability to remember. it is a test of what the runtime sends.

that distinction makes the result easier to inspect. there is no need to guess whether an answer changed because of sampling. i can open the saved request and check whether the codeword was there.

the lab stores the exact request before calling the provider, alongside the event numbers used to construct it. the browser console shows that request next to the complete event history. changing the selection leaves the earlier events intact.

the selection also keeps tool exchanges together. a tool call and its result belong to the same completed turn; slicing through the middle would produce a different kind of problem. the record needs enough structure to choose a coherent piece of the conversation.

this is a small mechanism, but it gives a useful place to start debugging. if an agent appears to have lost something, inspect the request. was the information selected? did it arrive in a usable form? the saved transcript alone cannot answer those questions.

the next experiment can ask which pages deserve a place on the desk: recent turns, a summary, retrieved details, or some combination. each choice should leave a record of what it included.

keeping the archive and arranging the desk are separate jobs. both need to be deliberate.

inspect the two saved requests