Large language models (LLMs) have shown strong capabilities in financial analysis and reasoning, spurring the development of agent‑based trading frameworks. Existing approaches either focus on long‑horizon forecasting or operate as stateless analyzers, limiting their applicability in complex trading environments.\ \ To fill this gap, we introduce META (Memory Enhanced Trading Agent), the first RAG‑style episodic‑memory‑augmented multi‑agent system for financial decision‑making. META consists of three core components:\
- Indicator agents: a family of specialized technical‑indicator agents (Trend, MACD, Stochastic, RSI, SMA, AVWAP, Heikin‑Ashi, etc.) that compute and report their respective signals.\
- Decision agent: aggregates the reports, adaptively re‑weights the signals according to the current market state, and outputs buy/sell actions.\
- Memory module: encodes each trading episode as a market‑state embedding together with outcomes and reflections. By retrieving past episodes with similar embeddings, the module supplies relevant experiences and dynamically adjusts indicator weights for the new episode.\ \ Empirical results show that META achieves higher directional accuracy and greater robustness in short‑horizon evaluations. The study demonstrates that episodic memory offers a powerful regime‑aware, interpretable, and low‑latency mechanism for trading decisions. The code is released on GitHub.\ \ Review: META cleverly integrates memory retrieval with multi‑indicator fusion, enabling the trading agent to self‑adapt across different market regimes. This boosts decision quality while preserving real‑time performance, marking a significant advance in intelligent financial agents.