Description:
The LLM does not retain user-provided contextual information (specifically the user's name) across turns within the same chat session. Initially, when prompted with a question such as “What is my name?”, the model correctly responds that it doesn't know. However, even after explicitly providing the name (e.g., “My name is Vikash”), subsequent prompts such as “What is my name now?” still result in the model responding with no knowledge of the name.

Model in question: granite-3.1-8b-lab-v2.1

Expected Behavior:
The LLM should utilize the conversational context window to reference recent user inputs. After a user provides their name in-session, the model should be able to infer or recall it in follow-up messages, assuming the context window is not exceeded.

[fzatlouk@localhost ~]$ nvidia-smi
Mon Jun 9 22:25:09 2025
-----------------------------------------------------------------------------------------

---------------------------------------{}--------------------{}++{}--------------------

-----------------------------------------------------------------------------------------

-----------------------------------------------------------------------------------------

atlouk@localhost ~]$ ilab model list

image:
image: registry.stage.redhat.io/rhelai1/bootc-nvidia-rhel9:1.5.1-1749187229
transport: registry
version: 9.20250429.0
timestamp: null
imageDigest: sha256:0acab414c92b8f1da20553a6c0041df3fb438bc3d74b331f51a1de2816fb7218