How PayEcho Forced Its LLM to Act on Customer History, Not Ignore It

Engineers at PayEcho, a payment-recovery and credit-decision platform, found that integrating recalled customer history into an LLM's context was not enough to change its recommendations. The model consistently produced generic advice even when months of behavioral data were available, treating historical information as background rather than evidence. The team solved this by requiring the model to explicitly cite specific past outcomes — such as a customer ignoring emails but responding to WhatsApp — as justification for each recommendation. Retrieval and generation were kept as separate pipeline stages, making it easier to diagnose whether a generic output stemmed from poor memory recall or failure to reason over retrieved data. Actual outcomes from each interaction are written back to memory, creating a feedback loop where past results become evidence for future decisions.