FinCon Trading Agents with Layered Memory and Verbal Feedback
Summary
The article describes a trading-system implementation inspired by FinCon, a multi-agent framework in which specialized analyst agents process market information and a manager combines their proposals, applies risk controls, and refines the strategy. Its design includes real-time risk management alongside review of completed episodes, plus conceptual verbal reinforcement feedback to evaluate agent contributions and trading decisions. Three memory types support the process: working memory for current tasks, procedural memory for reusable methods, and episodic memory for past events and outcomes.
The practical section builds an MQL5 analyst-agent module with trainable task queries, cross-attention for selecting relevant inputs, and standardized tensor outputs representing proposed actions. In this adaptation, each agent has its own memory. The article reports a test with 47 trades and a 42% win rate; most balance growth came from one profitable trade, while the balance otherwise stayed in a narrow range. These results suggest unstable performance and a need for further refinement. The account does not establish generalization or durable profitability, and the reported test alone is limited evidence for the framework’s effectiveness.
Key ideas
- A manager agent aggregates specialized analysts’ proposals and performs risk control and strategy refinement.
- The design combines immediate risk management with learning from completed episodes through conceptual verbal feedback.
- Working, procedural, and episodic memory serve different roles in current processing, reusable methods, and historical experience.
- Analyst agents use trainable queries and cross-attention, then express proposed decisions in a shared output format.
- The reported test had a low win rate and relied heavily on one profitable trade, indicating that results were not stable.
Tags
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.