Agent Memory with Episodic Retrieval for Financial Decision-Making

Nuoyue Xu; Jiang Liu; Wenxuan Huang; Xiang Zhang; Juntai Cao; Jiaqi Wei

cs.AISep 23, 2026

Agent Memory with Episodic Retrieval for Financial Decision-Making

Authors: Nuoyue Xu, Jiang Liu, Wenxuan Huang, Xiang Zhang, Juntai Cao, Jiaqi Wei

Organizations: University of British Columbia · University of Science and Technology of China · Fudan University · Zhejiang University

Abstract

Large language models (LLMs) have demonstrated strong capabilities in financial analysis and reasoning, inspiring recent advances in agent-based trading frameworks. While these systems show promise, prior approaches either emphasize long-horizon forecasting or operate as stateless analyzers, limiting their applicability to the demands of trading in complicated settings. To address these gaps, we introduce META (Memory Enhanced Trading Agent), the first RAG-like episodic-memory-augmented multi-agent framework for financial decision making. META integrates a family of specialized indicator agents (e.g., Trend, MACD, Stochastic, RSI, SMA, AVWAP, Heikin-Ashi) with a Decision Agent that fuses their reports, and a Memory module that retrieves and updates past trading episodes encoded as market state embeddings with outcomes and reflections. By recalling relevant experiences and adaptively reweighting signals under similar market regimes, META achieves improved directional accuracy and robustness under short-horizon evaluation. Our results demonstrate that episodic memory provides a powerful mechanism for regime-aware, interpretable, and low-latency decision-making in trading and decision making. The code of this project is released on GitHub.

Figures & tables

Figure 1: Architecture of the META memory system. The framework consists of three core components: Memory System Establishment, Relevant Memory Retrieval, and Memory Database Management.

Figure 2: The META cognitive workflow, illustrating the interplay of its three core operators. The process unfolds in stages: (1) Perception ( $\Phi$ ) : A distributed ensemble of Signal Agents encodes the market state in parallel. (2) Synthesis ( $\Pi$ ) : Real-time reports are fused with retrieved historical precedents to formulate a decision. (3) Memory ( $\Psi$ ) : The memory buffer is updated via a reflection-augmented consolidation cycle.

CL
Method	Accuracy $\alpha$ (% $\uparrow$ )	$\Delta\alpha$ (% $\uparrow$ )	$R_{cc}$ $\uparrow$	$R_{\max}$ $\uparrow$	$R_{\min}$ $\uparrow$	$R_{\text{sim}}$ $\uparrow$
+ Random	49.0	–	-0.312	0.999	-1.384	-0.255
+ QuantAgent	57.7	+17.8%	-0.195	1.181	-1.202	-0.133
+ META (Ours)	56.0	+6.1%	-0.484	1.024	-1.187	-0.433
ES
+ Random	41.3	–	0.006	0.560	-0.539	0.006

Table 1: Performance comparison across assets. Accuracy (

\alpha

), relative improvement (

\Delta\alpha

), correlation

R_{cc}

, maximum and minimum returns (

R_{\max}

R_{\min}

), and similarity metric (

R_{\text{sim}}

). Red indicates better performance, Green indicates worse, and black numbers denote baseline or neutral results.

Panel of Figure 3 — Figure 3: Similarity (top row per asset) and variance (bottom row) distributions for text embeddings vs. indicator vectors across all assets.

Figure 4: Accuracy comparison of text-embedding vs. indicator-vector retrieval.

Asset	Similarity (Mean)		Variance (Mean)
	Text	Indic.	Text	Indic.
BTC	0.9002	0.6428	0.0002	0.0031
CL	0.9089	0.6905	0.0002	0.0028
ES	0.9036	0.6853	0.0003	0.0030
NQ	0.9139	0.6826	0.0002	0.0033
QQQ	0.9095	0.7011	0.0003	0.0026

Table 2: Comparison of mean similarity and variance between text embeddings and indicator vectors. Indicator vectors show lower similarity but higher variance, leading to more discriminative retrieval.

Panel of Figure 5 — Figure 5: Top- $k$ and threshold retrieval analysis. For each asset: # memories retrieved ( $\tau=0.7$ ), average top- $k$ similarities ( $k=10$ ), and heatmap of retrieval counts over $k$ and $\tau$ .

Panel of Figure 6 — Figure 6: Accuracy versus number of memories. Each curve shows the moving average of directional accuracy as META’s memory base expands. The analysis begins at 300 memories, corresponding to the point where episodic retrieval starts to meaningfully influence decision-making.

Asset	Text Acc.	Indic. Acc.	Text Time	Indic. Time
CL	50.0	52.0	3.11	0.00610
ES	56.0	64.0	3.18	0.00580
NQ	60.0	62.0	3.15	0.00595
QQQ	50.0	61.0	3.12	0.00620
BTC	50.0	56.0	3.17	0.00575

Table 3: Accuracy (%) and retrieval latency (s) comparison between text-embedding and indicator-vector memory retrieval.

Figure 7: Performance of META under different market conditions across assets.

Asset	Method	Accuracy $\alpha$ (% ↑)	$\Delta\alpha$ (↑)
CL	w/o Mem	52.5	–
	w/ Mem	56.0	+3.5
ES	w/o Mem	50.0	–
	w/ Mem	64.0	+14.0
NQ	w/o Mem	40.0	–
	w/ Mem	62.0	+22.0

Table 4: Ablation study of memory system.

Figure 8: Impact of removing each technical indicator on META’s directional accuracy. Larger decreases indicate greater importance of that indicator to overall system stability and predictive accuracy.

Explore similar work

Learning to Trade Like an Expert: Cognitive Fine-Tuning for Stable Financial Reasoning in Language Models
Apr 18, 2026Yuchen Pan, Soung Chang LiewFrontier Models Multiple-Choice Questions
Department of Information Engineering, The Chinese University of Hong Kong
PandaAI: A Practical Agent CQ2 for Neuro-symbolic Data Analysis And Integrated Decision-Making in Quantitative Finance
Jun 5, 2026Yuqi Li, Siyuan Liu, Bingjun LiuAlgorithmic Trading Neuro-Symbolic Framework
Panda AI
Agentic Trading: When LLM Agents Meet Financial Markets
May 19, 2026Yihan Xia, Panpan You, Taotao Wang +4Large Language Model Agents Evaluation Protocols
College of Electronic and Information Engineering, Shenzhen University, Shenzhen, China

Apr 18, 2026, cs.LGJ/K move · Enter open · S save

Learning to Trade Like an Expert: Cognitive Fine-Tuning for Stable Financial Reasoning in Language Models

Yuchen Pan, Soung Chang Liew

Department of Information Engineering, The Chinese University of Hong Kong

Frontier Models Multiple-Choice Questions Experts

Open paper AskPDF arXiv

Recent deployments of large language models (LLMs) as autonomous trading agents raise questions about whether financial decision-making competence generalizes beyond specific market patterns and how it should be trained and evaluated in noisy markets lacking ground truth. We propose a structured framework for training and evaluating such models. Central to our approach is a curated, multiple-choice question (MCQ) dataset derived from classic textbooks and historical markets, verified by an AI committee, enriched with structured reasoning traces, and augmented to reduce shortcut learning. To evaluate whether performance on isolated MCQs generalizes to real-world trading, we introduce a two-stage protocol combining test-set evaluation with an MCQ-based chronological trading simulation. Extensive evaluations across market regimes provide statistically robust evidence that open models trained with our framework exhibit competitive, risk-aware behavior over time, outperform open-source baselines, and approach frontier-model performance at smaller scale. We release the dataset and evaluation framework to support further research.

Abstract

Explore similar work

Learning to Trade Like an Expert: Cognitive Fine-Tuning for Stable Financial Reasoning in Language Models

PandaAI: A Practical Agent CQ2 for Neuro-symbolic Data Analysis And Integrated Decision-Making in Quantitative Finance

Agentic Trading: When LLM Agents Meet Financial Markets