session_id to show that sessions do NOT carry memory; (2) a manual history loop that does; (3) node_input_overrides to swap the LLM system prompt at runtime.
You’ll end up with. Printed results from each run showing that session_id is for tracing only, while client-side history (or node_input_overrides) is what actually changes LLM behaviour.
Expected output
session_id is a tracing/analytics tag — it groups runs in the dashboard but doesn’t feed prior messages back into the LLM. For conversational memory, pass the running transcript as the prompt yourself, or use a Chatbot.
See also
Intermediate results
Stream every node’s output as it runs.
Background runs
Fire a run and poll for completion.
Pipeline reference
Full method surface.
