Title: Managing Complexity: An Interaction-Native Knowledge Framework for Financial LLM Agents

Abstract:

A primary reason for the failure of financial AI agents is their tendency to offload complexity onto the user. In current setups, users are forced to repeatedly articulate their goals, risk tolerances, portfolio details, historical judgments, and evolving market assumptions. Meanwhile, the agent operates on a cycle of answering, retrieving, acting, and forgetting. In the financial sector, this dynamic is more than an inconvenience; it introduces significant risks. During critical tasks such as market analysis, copy-trading review, and trade preparation, lost context and outdated memory can lead to operational latency, recurring errors, poor auditability, and potentially unsafe decision-making.

To address this, we introduce the Interaction-Native Knowledge Harness (InKH), an architectural framework designed for financial Large Language Model (LLM) agents that internalizes complexity rather than shifting it to the user. InKH transforms events related to users, markets, portfolios, and tools into structured operational knowledge. The system employs passive knowledge injection to construct a bounded working context buffer prior to the main model execution. Additionally, it utilizes temporal graph memory for rapid retrieval, a wiki-based audit surface to ensure human-readable governance, and background extraction processes that incorporate maturity, decay, and write-time invalidation.

We assessed InKH using a reproducible, controlled synthetic benchmark. The evaluation involved 24 random seeds, 4 rounds, 80 episodes per round, and 6 baseline models, resulting in 46,080 baseline-conditioned evaluations. InKH demonstrated a mean task quality score of 0.815 with a latency of 900 milliseconds. When compared to an agent-driven wiki-walk memory system, InKH reduced latency by 82.95 percent, lowered token costs by 82.29 percent, and decreased the usage of stale knowledge by 96.58 percent. Simultaneously, it improved task quality by 0.108 and enhanced traceability by 0.461.

Furthermore, compared to a temporal-graph system lacking invalidation mechanisms, InKH increased quality by 0.050 and cut stale-memory usage by 96.58 percent, all while maintaining comparable serving costs. These findings reinforce a central design thesis for financial AI: widespread adoption is contingent upon the system absorbing complexity instead of transferring it to the user. It is important to note that this benchmark validates architecture-level behavior and does not reflect live trading performance.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC