FinMem: Layered Memory and Risk Profiles for an LLM Trading Agent
Summary
The article describes FinMem, an LLM-based trading-agent design organized around profiling, memory, and decision-making. Its profile combines company and sector context with a selected risk style: risk-seeking, risk-averse, or self-adaptive. The memory design pairs a working space for summarizing, observing, and reflecting with long-term layers that retain events according to their relevance and expected rate of decay. For a query, the agent retrieves ranked memories and combines them with market information to produce a direction, rationale, and influential events.
The account distinguishes training, when future price changes are available as labels, from testing, when the agent must act without future prices. It also describes short-term and extended reflections that can feed back into deeper memory. The article reports that the cited research found favorable results against other autonomous models, but provides no detailed metrics or experimental controls in the excerpt. Its implementation discussion is a work in progress and explicitly excludes an LLM, so the described architecture and claims should not be treated as independently validated trading performance.
Key ideas
- FinMem separates profiling, layered memory, and decision-making into distinct components.
- Long-term memory layers retain information according to relevance and temporal decay, while working memory summarizes and reflects on current inputs.
- Retrieved events are ranked by novelty, relevance, and importance before being combined with market observations.
- The agent's risk profile can change in response to recent cumulative returns.
- Training uses future price changes as direction labels, while testing removes access to future prices; the article offers limited detail for evaluating reported results.
Tags
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.