Leading AI labs adopt selective memory for fintech agents

Anthropic’s engineering team recently removed more than 80% of the system prompt from the latest Claude Code model generation and reported no measurable decline in coding performance. The engineers published their findings under the term progressive disclosure.

Progressive disclosure means loading tools and context on demand instead of preloading long, static prompts that list many rules and references. Anthropic’s engineers reported that models performed as well or better when given only the context relevant to the specific task at the moment.

Ternary Capital Research has developed a parallel design for financial agents called DeepCQ. DeepCQ uses a multi-tier memory architecture that includes live application state, a session cache, dialogue history, persistent analytical notes and longer-term records. The system scores sections of that architecture and returns only the snippets deemed relevant to the current query.

Ternary’s implementation does not use embeddings or vector databases. The company states that avoiding vector infrastructure simplifies deployment in environments with strict data-handling and storage rules.

Anthropic and Ternary note specific operational effects for banks and asset managers. Agent contexts in those institutions often include compliance text, past case notes and regulatory references that are rarely needed for a single user query. Sending all that material on every request increases latency, token use and the amount of data exposed to the model, the two teams report.

Both Anthropic and Ternary recommend treating memory and retrieval as a core architectural layer in production agent systems. Their documentation and technical notes state that selective retrieval reduces token volume and shortens response times at scale.

Anthropic and Ternary reached similar technical conclusions independently about supplying context on demand for agent workflows in regulated financial environments.