Multimodal Agent MemorySubmit a paper

FinAgent

A Multimodal Foundation Agent for Financial Trading: Tool-Augmented, Diversified, and Generalist

Wentao Zhang, Lingxuan Zhao, Haochong Xia, Shuo Sun, Jiaze Sun, Molei Qin, et al. · 2024-02-28

Read the paperPDF

Management
T2 Dynamic Memory
Memory changes during deployment through a stable management policy.
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. memory module holding market-intelligence summaries plus low-level and high-level reflection analyses, embedded with an LLM and matched by vector similarity

    1. Represented asPerception-to-text records & summariesCross-represented explicit
    2. Stored inVector-indexed collection
    3. Retrieved bySimilarity search
    4. Source evidenceUnclear
    Sec. 4.2 states that all analyses and summaries produced by the market-intelligence, low-level-reflection and high-level-reflection modules are stored in the memory module, which uses a vector storage architecture. Sec. 4.1 adds a dedicated query-text field to each market-intelligence output and embeds it with an LLM so that past entries with similar content are retrieved by vector similarity. The stored items are natural-language records distilled from numeric, textual and chart inputs; the paper does not state that the Kline chart images themselves are kept or linked from the memory entries, hence source accessibility is left unclear.Full paper: 4.1 Market Intelligence Module; 4.2 Memory Module; 4.3 Reflection Module · arXiv HTML v3, Sec. 4.1-4.3 · Figure 2 (overall architecture)

How memory changes (T2 operations)

  • Update · Rule-based update

    At every trading step of the deployed decision loop the newly produced market-intelligence summary and the new low-/high-level reflections are written into the memory module and become retrievable for subsequent trading days; the memory content therefore grows and is reused during inference under a fixed write rule, with no change to the write or retrieval policy itself.

Related systems