At the Portfolio Banking Technology 2026 conference, Schäffer Krisztián, AI Strategy Lead at ARDINSYS, argued that simply ingesting a firm's document corpus is not enough for reliable AI agents in banks. He said the key is to build a multi-layered, well-structured and dynamically updated organizational memory.
Practical limits of LLMs and agents
Schäffer noted that language models (LLMs) are, in effect, stateless mathematical functions: they do not retain information across consecutive calls. Consequently, conversational history must be resent to the model on every call, which means confidential data needed for decision-making must be transmitted to the servers serving the model.
Agents operate by having the model perform tool calls and effectively converse with itself, accumulating information in a context window "notebook" until it can produce a final answer. Because the context length grows with each step, agents' resource needs and costs increase — Schäffer described this growth as proportional to the square of the number of steps — and caching techniques can only partially mitigate that increase.
Why loading documents is insufficient
Corporate documents are heterogeneous, often incomplete or out of date. According to Schäffer, simply loading a company's document stock does not guarantee extraction of relevant, up-to-date information for AI assistants. Instead, an organizational memory that goes beyond static data and is refreshed dynamically based on user interactions is required.
ARDINSYS approach: Volumentor and memory layers
In the Volumentor work environment developed by ARDINSYS, the company organizes knowledge into multiple stacked layers:
- Base layer: searchable tables of static documents, source code and configurations.
- Dynamic layer: transcripts of user conversations are analyzed by independent "curator agents"; when new information or feedback is identified, it is automatically added to a central knowledge repository stored as Git-based Markdown files, while outdated items are archived.
- Associative layer: preprepared knowledge elements (described by Schäffer as "quick directional breadcrumbs") are inserted into the context at the start of a conversation to help the agent choose the correct direction.
Architecture independence and vendor risk
Schäffer stressed that base models have become utility-like, so preserving architecture independence and avoiding vendor lock-in is crucial when designing enterprise systems. This reduces risk when models, providers or infrastructure need to be changed.
Why this matters for banks
Given banks' particular concerns over confidential data handling, compliance and cost efficiency, solutions that rely only on static document stores are unlikely to be sufficient. A layered, continuously maintained organizational memory and a vendor-agnostic approach to models can reduce unnecessary data transfer, optimize costs and improve agent reliability.
Note
An AI assistant contributed to the preparation of the Portfolio article; the final content was edited and verified by the outlet's journalist.
