The Need for Personalized Memory in Financial Agents

Existing benchmarks for LLM agents often fail to capture the nuances of long-term, personalized memory—a critical requirement for financial applications where agents must track individual user history, preferences, and specific life events over time. FinPerMA (Financial Personalized-Memory Agent) addresses this gap by grounding agent evaluation in psychological theories of memory, specifically focusing on how agents retrieve and synthesize information from a user's unique financial timeline.

Benchmark Architecture and Evaluation

FinPerMA moves beyond simple retrieval tasks by requiring agents to demonstrate 'event-grounded' reasoning. The benchmark tests an agent's ability to:

• Maintain Contextual Continuity: Track financial events (e.g., income changes, major purchases, investment shifts) across long interaction histories.
• Apply Personalized Logic: Use stored user-specific data to provide tailored financial advice rather than generic, one-size-fits-all responses.
• Handle Memory Decay and Updates: Correctly update the agent's internal 'state' when new information contradicts or updates previous user data.

By grounding the benchmark in theory, the researchers aim to move the field toward agents that function more like long-term financial assistants rather than stateless query-response engines. The benchmark provides a structured framework for measuring how effectively an agent can navigate the trade-offs between memory capacity, retrieval accuracy, and the synthesis of personalized financial narratives.