Companion agents fail when memory is a flat chat log. I split facts goals and tone into separate stores so one bad turn does not rewrite the whole persona.
Exactly. When everything is stored in one large block of context, a noisy signal can have too much influence on the agent’s reasoning. Separating facts, relationships, and task-specific information gives the model a clearer and more stable way to work.
Puneet Khandelwal
Spot on. Dumping raw error logs into a prompt is just a lazy shortcut. We need structured error context instead.