How a Financial AI Agent Was Fixed to Hold One Belief Per User Fact

A developer building an AI financial advisory agent discovered that storing memory by session caused contradictory user beliefs — such as both 'cautious' and 'aggressive' risk profiles — to surface simultaneously during retrieval. The flaw stemmed from inserting new rows each session with no concept of a 'current' belief, leaving downstream agents unable to determine which version was valid. Two additional bugs were also uncovered: the agent fabricated financial data it was never given, and the system performed keyword matching despite being assumed to use semantic search. The team redesigned the memory layer around a 'one belief per fact' principle, where each fact has a stable key, a single current value, and a versioned history for auditability. Embeddings are used solely for retrieval ranking, while stable keys handle identity and metadata handles governance, preventing corrupted or conflicting records.