Why Today's AI in Finance Forgets

Financial analysis using artificial intelligence isn't just about reading numbers. It simultaneously requires calculating, searching for information, interpreting charts and tables, and temporal anchoring — all across different types of documents. Annual reports, stock charts, accounting spreadsheets, press releases — each source speaks a different language and AI must correctly connect them.

The problem is that current models — including those built into tools like ChatGPT, Claude, or Gemini — operate without memory across tasks. Every new question means starting over from scratch. They repeatedly discover the same reasoning strategies and fall into the same traps. In the world of finance, where a mistake can mean a bad investment decision, this is a fundamental weakness.

FinAcumen: Memory as the Key to Accuracy

The FinAcumen framework, created by a team from Beijing University of Posts and Telecommunications and Queen Mary University of London, solves this problem in a fundamentally new way. Its core is a module called Financial Memory (FM) — a persistent memory bank that stores experiences from previous tasks.

But this isn't about simply storing everything the model has ever done. FinAcumen rigorously separates successful strategies from cautionary lessons learned from mistakes. When the model faces a new question, it searches its memory and finds relevant experiences only if their semantic similarity to the current task exceeds a calibrated threshold. If no relevant experience exists, the model explicitly falls back to its base reasoning — thus preventing irrelevant "memories" from degrading answer quality.

This selective memory activation mechanism is crucial for finance. As the authors state in the published study: "In high-stakes environments such as finance, irrelevant retrieval can directly degrade reasoning quality, and hallucinated evidence can lead to incorrect analytical conclusions."

Four Tools That Keep the Model Grounded

In addition to memory, FinAcumen also deploys a set of deterministic financial tools (Financial Tools, FT):

• Numerical Reasoning Engine — a specialized computation module for financial mathematics that eliminates arithmetic errors common in language models
• Grounded Data Retrieval — targeted retrieval from financial documents with an emphasis on verifiable sources
• Visual Structure Decoder — reading charts, tables, and diagrams typical of financial reporting
• Answer Consolidation Gate — final answer verification that checks consistency across all used sources

Thanks to this combination of tools and memory, FinAcumen doesn't rely solely on the model's "intelligence" but provides it with solid ground under its feet in the form of verified calculations and verifiable data.

Results: A Small Model with Memory Beats Specialists

The researchers tested FinAcumen on four specialized financial benchmarks: SEC-NUM (quantitative reasoning from annual reports), FinMMR (numerical reasoning from charts), FinTMMBench (temporal multimodal retrieval), and FinMME (general financial multimodal evaluation).

The results are remarkable. FinAcumen, built on a frozen 8B vision-language model (meaning a model that was not fine-tuned on financial data at all), consistently outperforms specialized financial models like Fin-R1, Open-FinLLMs, and LLM Pro Finance Suite. In many tests, it approaches the performance of top proprietary models — and at a fraction of their size.

A key insight from the sensitivity analysis: the more relevant experiences the model accumulates, the more stable its performance — especially under conditions where retrieval is inaccurate or documents contain noise. In other words: memory serves as protection against failure.

What This Means for Czech Companies and Investors

For the Czech financial sector — whether banks, investment funds, or fintech startups — FinAcumen brings an interesting perspective. It shows that building reliable financial AI doesn't require training giant models from scratch. A much more efficient path may be to equip an existing model with good memory and reliable tools.

This is especially important in the European context, where the AI Act emphasizes transparency and verifiability of AI system decision-making. A framework that explicitly separates successful strategies from mistakes and allows tracing why it arrived at a given conclusion is exactly the direction financial AI should take.

The open-source nature of FinAcumen (code is available at anonymous.4open.science) also means that Czech developers and companies can freely test and adapt the framework for their needs — for example, for analyzing annual reports of Czech companies or for automated processing of financial documents in Czech.

Limits and What's Next

The authors themselves acknowledge that the current version has its limits. The memory bank currently doesn't work with the concept of "forgetting" — it could become overloaded over time. There is also a lack of support for multilingual financial documents, which is significant for the European market with its many languages. And the model works with static documents, not with real-time market data.

Nevertheless, the direction FinAcumen charts is clear: the next generation of AI for finance will not be built on larger models, but on smarter management of experience. Just as a human analyst grows with every completed task, AI agents must also learn from their own history.