Multimodal Agent MemorySubmit a paper

FinAcumen (Financial Memory + Financial Tools harness)

FinAcumen: Financial Multimodal Reasoning via Self-Evolving Experience Memory Harness

Pianran Guo, Pengcheng Zhou, Yuchen Jian, Shuhua Chen, Zhongliang Yang, Linna Zhou · 2026-08-23

Read the paperPDF

Management
T1 Static Memory
Memory is built once and reused without change during deployment.
Primary modality
Image
Also: text
Representation
Cross-represented explicit

How experience becomes memory

Each path is one memory component, traced from what is stored to how it is found again.

  1. Financial Memory (FM) bank of distilled Findings and Cautions entries harvested from multimodal financial reasoning trajectories

    1. Represented asPerception-to-text records & summariesCross-represented explicit
    2. Stored inVector-indexed collection
    3. Retrieved bySimilarity search
    4. Source evidenceSource not retained
    Each FM entry holds the original question, the gold answer, distilled 'Findings' (successful patterns) and 'Cautions' (failure-derived guard rules), plus a pre-computed text embedding. Entries are produced by a summary agent that scores and summarises sampled trajectories over chart/table/text financial problems (Eq. 2). Retrieval embeds the incoming query with the shared encoder, keeps entries whose cosine similarity exceeds tau = 0.65 (Eq. 3-4), then deduplicates and injects the top-k into the prompt. No chart image or table artifact is stored in the entry, so the visual source is not recoverable from memory.Full paper: Sec. 3 (framework), Sec. 4.1-4.2 (memory construction and selective retrieval), Appendix B (Algorithm 1) · HTML Sec. 3-4; Appendix B · Eq. 2-4; Algorithm 1; Figure 3 (bank saturation ~1800 entries)

Related systems