Source-linked AI summary
Absorbing Complexity: An Interaction-Native Knowledge Harness for Financial LLM Agents
Ailiya Borjigin, Igor Stadnyk, Ben Bilski, Maksym Chikita, Dmytro Kyrylenko, Sofiia Pidturkina, Julia Stadnyk
TL;DR
Financial AI agents often make users repeatedly manage fragmented context and changing assumptions. InKH absorbs that complexity through passive context injection, structured memory, and a wiki audit surface; a controlled synthetic benchmark reports improved quality, latency, stale-memory suppression, and traceability, but does not establish live trading profitability.
Problem
Financial agents often lose useful workflow context, requiring users to repeatedly coordinate goals, risk preferences, portfolio information, historical judgments, and changing market assumptions.
Method
InKH is an interaction-native architecture that continuously converts interaction traces into governed knowledge and injects relevant information into a bounded working context.
Results
InKH improves quality, lowers latency, reduces stale-memory usage, and increases decision traceability in a reproducible controlled benchmark relative to memory and non-memory baselines.
Takeaways & Limitations
The results support interaction-native, governed knowledge harnesses as a systems target for financial agents rather than turn-based retrieval with disposable context.
Takeaways & Limitations
The evaluation uses a controlled synthetic benchmark with simulator-defined quality, abstractly simulated graph-backed retrieval, and no reported public-data replay or live trading profitability.
Abstract
from arXiv · showhide
Financial AI agents often fail for a simple reason: they make users carry the complexity. A user must repeatedly restate goals, risk preferences, portfolio context, past judgments, and shifting market assumptions, while the agent answers, retrieves, acts, and forgets. In finance, this is not just inconvenient. In tasks such as market analysis, copy-trading review, and trade preparation, forgotten context and stale memory can create latency, repeated errors, weak auditability, and unsafe decisions. We propose the interaction-native knowledge harness (InKH), an architecture for financial LLM agents that absorbs complexity into the system. InKH converts user, market, portfolio, and tool events into structured operational knowledge. It uses passive knowledge injection to assemble a bounded working context buffer before the main model step, temporal graph memory for low-latency retrieval, a wiki audit surface for human-readable governance, and background extraction with maturity, decay, and write-time invalidation. We evaluate InKH on a reproducible controlled synthetic benchmark with 24 random seeds, 4 rounds, 80 episodes per round, and 6 baselines, producing 46,080 baseline-conditioned evaluations. InKH achieves mean task quality of 0.815 at 900 ms latency. Compared with agent-driven wiki-walk memory, it reduces latency by 82.95 percent, token cost by 82.29 percent, and stale-knowledge usage by 96.58 percent, while improving quality by 0.108 and traceability by 0.461. Compared with a temporal-graph system without invalidation, it improves quality by 0.050 and reduces stale-memory usage by 96.58 percent with comparable serving cost. The results support a design thesis for financial AI: adoption happens when complexity is absorbed by the system rather than transferred to the user. The benchmark validates architecture-level behavior, not live trading performance.
1 Introduction
Financial LLM agents often leave users to manage persistent financial context, while InKH proposes absorbing that complexity through structured, continuous knowledge management. The architecture combines passive context preparation, governed memory, and auditability, and is evaluated against memory and non-memory baselines.
- Turn-based financial agents lose useful context, causing recurring latency, duplicated reasoning, fragile personalization, and repeated rediscovery of risks.
- InKH treats financial cognition as continuous and converts interaction traces into structured, persistent, operational knowledge.The approach reduces the need for users to coordinate information, historical judgments, risk preferences, and changing market assumptions.
- InKH is positioned as complementary to execution-layer safety: safe financial agents require both upstream cognition management and downstream action controls.
- InKH combines passive knowledge injection, a bounded working context buffer, temporal graph memory, and a wiki audit surface.
- The architecture formalizes state, knowledge objects, retrieval utility, injection, decay, invalidation, maturity transitions, and governance constraints.
- The paper provides algorithms and implementation guidance for passive injection, background extraction, maintenance, graph-plus-wiki systems, entity matching, upserts, and latency budgeting.
- The study reports a reproducible controlled financial benchmark showing improvements in quality, latency, stale-memory suppression, repeated error reduction, and traceability.
2 Literature Review
Prior work spans continuous interaction, multi-tier and graph-backed memory, trajectory learning, retrieval, and financial-agent benchmarking. InKH builds on these strands while shifting retrieval from an agent-only query-time decision toward injected bounded working state.
- Continuous-interaction research motivates stateful financial cognition, although InKH adopts the architectural lesson rather than the cited multimodal training recipe.
- Memory research includes wiki-maintained knowledge, multi-tier memory management, production long-term memory, dynamically organized agentic memory, and temporal knowledge graphs.
- Trajectory-learning studies show that agents can improve without finetuning through reflection, reusable knowledge, skills, memory, iterative refinement, adaptive retrieval, and self-critique.
- Finance-specific agent research remains comparatively sparse, while newer benchmarks cover financial QA, SEC-filings research, business-driven tasks, and agentic retrieval.
- Financial-agent execution research addresses survivability, constrained execution, action shielding, and auditable compliance; InKH extends this trajectory upstream by maintaining evolving financial state.
- Unlike conventional retrieval and tool-use approaches, InKH makes relevant knowledge injection part of bounded working state rather than solely an agent query-time decision.
3 Problem Formulation
The paper models a financial agent as an event-driven system that maintains user, market, risk, workflow, and temporal-graph state. Its problem formulation centers on assembling bounded, governed context while balancing task quality, cost, action risk, knowledge reuse, and memory maintenance.
- 3 Problem Formulation: Financial-agent events include user turns, tool observations, market updates, portfolio changes, and internally generated risk signals.
- State and knowledge objects: The agent state combines user, market, risk, workflow, and temporal knowledge-graph state.
- State and knowledge objects: Knowledge objects carry type, scope, content, provenance, confidence, maturity, regime, validation, and invalidation metadata.
- Working context buffer: The working context is assembled from active entities, intent, risk class, graph neighborhoods, and governance-admissible candidates under a token budget.
- Working context buffer: The system prepares context before the next reasoning step instead of requiring the model to search for it.
- Objective: The objective formalizes a tradeoff among answer quality, serving cost, action risk, and future knowledge benefit.
- Knowledge update, decay, invalidation, and maturity: Background extraction updates the graph, while decay, regime changes, validation, reuse, downstream utility, human review, and contradiction-triggered invalidation govern knowledge maintenance.
- Theoretical propositions: Passive injection is cheaper when retrieval overhead exceeds its irrelevant-context penalty, while higher action risk requires higher acceptable knowledge maturity.
4 Interaction-Native Knowledge Harness
InKH separates low-latency online retrieval from offline workflow extraction and human-readable audit synthesis. The architecture uses a temporal graph for compact governed context and a wiki for inspection, review, and maintenance.
- Online and offline paths: The online path retrieves a governed temporal-graph neighborhood and injects it into the working buffer before the main agent step.
- Online and offline paths: The offline path extracts completed workflows, upserts knowledge, invalidates contradicted items, and summarizes results into a wiki audit surface.
- Retrieval and audit separation: The temporal graph stores canonical entities, typed relations, provenance, validation times, and invalidation metadata for compact retrieval.
- Retrieval and audit separation: The wiki stores readable asset, trader, strategy, risk, and maintenance pages for inspectability and human review.
- Baseline comparison: The baselines are compared by which architectural components they contain.
- Engineering constraints: Production retrieval should be algorithmic and budgeted rather than requiring an LLM-driven wiki walk.
- Engineering constraints: Production systems should canonicalize aliases, invalidate superseded knowledge at write time, and retain the wiki primarily as an audit surface.
5 Experimental Design
The evaluation uses a controlled synthetic benchmark to test architecture-level behavior under explicit abstractions and assumptions. It compares six systems using quality, cost, latency, stale-knowledge, traceability, and reproducibility measures.
- Evaluation scope: The study reports only Stage A results from a controlled synthetic benchmark; a public-data replay extension is specified but not executed.Table 2 distinguishes the implemented evaluation from the planned replay extension.
- Benchmark design: The artifact evaluates 24 seeds, 4 rounds, and 80 episodes per round across market analysis, portfolio review, copy-trading evaluation, and trade preparation.This produces 7,680 workflows per baseline and 46,080 baseline-conditioned evaluations.
- Assumptions: The architecture-level study models baselines through characteristic token budgets, retrieval behavior, and latency distributions rather than one proprietary API.Token cost is modeled as $3.00 per million tokens plus $0.002 per tool call, while quality, stale-memory violations, and traceability use simulator-defined gold requirements.
- Scope: Reported results validate system behavior rather than live trading profitability.This scope boundary is an explicit assumption of the evaluation.
- Baselines: The six compared systems are ModelOnly, ToolAgent, SimpleMem, WikiWalk, Khnoinv, and full InKH.WikiWalk and InKH both use compiled persistent knowledge, but only InKH adds passive injection and write-time invalidation.
- Metrics and testing: The metrics include quality, latency, total tokens, stale-knowledge usage, decision traceability, and estimated cost efficiency defined as Q tokens/1000.Uncertainty uses 95% bootstrap intervals over seed-level means, and pairwise comparisons use paired Wilcoxon signed-rank tests over 24 seed-level means.
6 Results
InKH achieves the strongest combined quality and traceability while lowering latency and stale-memory usage against persistent-memory baselines. Its largest advantage appears after shocks, where invalidation supports reuse without simply storing more knowledge.
- Overall results: The full InKH baseline has the highest task quality and traceability among all systems, with lower latency than every nontrivial retrieval baseline.Figure 2 places InKH in the best region of the quality-latency frontier among persistent-memory baselines.
- Baseline comparisons: 82.95% lower latency, 82.29% lower token load, 0.108 higher quality, 96.58% lower stale-memory usage, and 0.461 higher traceability distinguish InKH from WikiWalk.These are the reported InKH-minus-WikiWalk comparisons.
- Statistical testing: All gains against the three most relevant memory baselines are statistically significant.Table 4 reports paired comparisons against those memory baselines.
- Shock adaptation: Only full InKH improves materially from Round 1 to Round 4; other baselines remain flat or regress after shock introduction.After shocks, stale-memory usage spikes in every memory baseline except the invalidation-enabled system.
- Task families: InKH is strongest across all four task families, with the largest gains on copy-trading evaluation and trade preparation.On high-risk workflows, InKH reaches quality 0.822, stale-memory usage 0.018, and traceability 0.999, versus 0.766, 0.336, and 0.923 for Khnoinv.
- Mechanism: InKH and Khnoinv accumulate essentially the same knowledge and maturity mass, so the observed difference is obsolete-memory invalidation rather than greater storage.The paper frames this as remembering under governance rather than merely remembering more.
7 Discussion and Limitations
The discussion recommends passive context assembly, graph retrieval, wiki auditability, and write-time invalidation as complementary product choices for financial agents. It also bounds the evidence to a controlled synthetic benchmark rather than live trading profitability.
- Product lessons: Passive injection should replace wiki walking in the foreground path because the system assembles context before model reasoning.WikiWalk and InKH both use compiled persistent knowledge, but only InKH combines passive injection with write-time invalidation.
- Product lessons: The graph should serve retrieval while the wiki serves audit and review, preserving interpretability without document-traversal costs.
- Product lessons: Invalidation, rather than additional memory volume, explains the performance gap between InKH and Khnoinv after outdated assumptions survive regime breaks.Both KH variants ingest the same amount of knowledge; only InKH invalidates obsolete items.
- Product lessons: The cognition and execution planes should be designed separately but coherently, pairing cognition survivability with execution survivability.The paper positions execution-layer safety work as downstream complement to its upstream cognition-plane contribution.
- Limitations: The evaluation is a controlled synthetic benchmark measuring architecture-level properties, not live trading profitability.The quality metric is simulator-defined, retrieval is abstractly simulated, and public-data replay is specified but not yet reported.
8 Conclusion
The conclusion presents InKH as a system that absorbs financial agents’ cognitive complexity through governed, continuously maintained knowledge. Its benchmark results support continuous cognition as one part of a broader architecture that also requires survivable execution.
- InKH absorbs cognitive complexity by maintaining structured state, injecting timely context, and transforming interaction traces into governed long-term knowledge.
- The reproducible benchmark shows improved quality, lower latency, sharply reduced stale-memory usage, and higher decision traceability than nontrivial memory baselines.The conclusion emphasizes governed remembering rather than merely storing more information.
- Future financial agents require both continuous cognition and survivable execution, addressed respectively by InKH and complementary execution-safety work.
A.1 Implementation defaults
The implementation appendix distinguishes architectural requirements from recommended starting values. Required items support the claimed architecture, while suggested defaults are implementation recommendations rather than optimality claims.
- Table 8 separates required architectural commitments from suggested defaults.“Required” denotes necessity for the claimed architecture; “Suggested” denotes a recommended starting value.
A.2 Synthetic benchmark mechanics
The benchmark artifact is a deterministic, seed-controlled simulator that computes quality from baseline priors, round effects, retrieval hits, missing requirements, stale use, and noise. It also provides reproducibility scripts and example raw, entity, and edge records.
- The released artifact implements the benchmark as a deterministic simulator with seed-controlled randomness.Its quality function is specified explicitly for reproducible evaluation.
- Q = clip(qb + βr(r −1) + βh · hits −βm · missing −βs · stale + ϵ) defines quality from baseline, round, retrieval, missingness, stale-use, and noise terms.The noise term follows a normal distribution with σ = 0.018.
- Missing gold requirements incur a penalty of 0.02 per miss, while stale-memory penalties are 0.11 for WikiWalk, SimpleMem, and Khnoinv and 0.04 for full InKH.Retrieval-hit bonuses and caps vary by baseline.
- Knowledge records: Raw evidence records store immutable source payloads with source type, reference, timestamp, trust tier, and content.
- Knowledge records: Entity records store canonical identifiers, aliases, entity types, summaries, and update timestamps.
- Knowledge records: Edge records connect entities to risks with relation types, evidence ids, confidence, maturity, regime tags, and validity intervals.
- Reproduction: Running python scripts/run_synthetic_suite.py regenerates workflow logs, summary tables, confidence intervals, paired tests, and result figures.