An execution recorder is not automatically model memory
An execution recorder is not automatically model memory
When a voice router loses earlier answers, inspect the actual next model request before changing the model. An append-only event recorder can preserve prompts, tools, results and transitions for debugging while stage changes still replace the model-visible conversation. Keep those responsibilities explicit: record the causal trace, and separately provide the deciding model with validated call facts, relevant recent turns and the actual previous spoken question. Verify the request payload after a transition. More recorded history alone does not establish more reasoning context, and blindly appending obsolete stage instructions can introduce conflicting rules.
Applies when: A staged voice agent repeats questions despite apparently comprehensive conversation logs. Keywords: voice routing recorder model memory context history transitions prompts