AI research wiki
agentic-systems · 2026-06-07 · Tier 2

InKH: Interaction-Native Knowledge Harness for Financial LLM Agents

InKH: Interaction-Native Knowledge Harness for Financial LLM Agents

Source: HuggingFace Daily Papers · arXiv 2606.01886 Raw: raw/huggingface/2026-06-07-absorbing-complexity-an-interaction-native-knowledge-harness.md

TL;DR

Financial AI agents fail because they make the user carry the complexity: the user restates goals, risk preferences, portfolio context, and shifting market assumptions while the agent answers and forgets. InKH (interaction-native knowledge harness) absorbs that complexity into the system. It converts user, market, portfolio, and tool events into structured operational knowledge, then uses passive knowledge injection to assemble a bounded working-context buffer before each model step, a 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. On a controlled synthetic benchmark (24 seeds, 80 episodes/round, 6 baselines, 46,080 evaluations), InKH reaches mean task quality 0.815 at 900ms latency — versus agent-driven wiki-walk memory, 82.95% less latency, 82.29% less token cost, 96.58% less stale-knowledge use, +0.108 quality, +0.461 traceability.

flowchart LR
  EV[User / market / portfolio /<br/>tool events] --> EX[Background extraction<br/>maturity, decay,<br/>write-time invalidation]
  EX --> TG[Temporal graph memory]
  TG --> INJ[Passive knowledge injection]
  INJ --> BUF[Bounded working-context<br/>buffer]
  BUF --> M[Main model step]
  TG --> WIKI[Wiki audit surface<br/>human-readable governance]
  M --> O[Quality 0.815 @ 900ms<br/>stale use -96.58%]
  STALE[Agent-driven<br/>wiki-walk memory] -.->|high latency,<br/>stale reads| W[Repeated errors,<br/>weak auditability]
  classDef input fill:#dbeafe,stroke:#3b82f6,color:#1e3a8a
  classDef stage fill:#e0e7ff,stroke:#6366f1,color:#312e81
  classDef output fill:#d1fae5,stroke:#10b981,color:#065f46
  classDef warn fill:#fee2e2,stroke:#ef4444,color:#7f1d1d
  class EV input
  class EX,TG,INJ,BUF,M stage
  class WIKI,O output
  class STALE,W warn

Key points

  • Write-time invalidation is the load-bearing piece. The big stale-memory reduction (96.58%) comes from invalidating knowledge at write time rather than hoping retrieval filters it later. Compared against a temporal-graph system without invalidation, InKH still improves quality by 0.050 and cuts stale use by the same margin at comparable cost — so invalidation, not the graph alone, drives the gain.
  • Passive injection vs agent-driven retrieval. Assembling a bounded context buffer before the model step (passive) beats letting the agent walk a wiki to fetch memory (active) on latency (82.95% lower) and token cost (82.29% lower) — the cost of agentic retrieval is the thing being removed.
  • Architecture validated, not live trading. The benchmark is a controlled synthetic environment; the paper is explicit that it validates architecture-level behavior, not market performance.

How this relates to prior wiki knowledge

  • Stale memory is the recurring failure mode. This directly addresses the diagnosis from the agent-memory cluster (05-15: STALE capped frontier models at 55.2% on implicit-conflict detection over long contexts) — stale, un-invalidated memory is what breaks long-running agents. InKH's write-time invalidation is a concrete mechanism for the "memory must be adaptive infrastructure, not a frozen RAG database" thesis.
  • Belief-clarity and faithful-compression cousins. Where MMPO (06-05) penalizes summaries that muddy belief and MemTrain (06-04) rewards faithful compression, InKH attacks the same staleness problem at the systems layer (invalidation + decay + maturity) rather than the training-objective layer. The two approaches are complementary: train memory to compress faithfully, then engineer the store to expire correctly.
  • The auditability angle. A human-readable wiki audit surface for governance is the finance-specific version of the traceability the responsible-ai page keeps asking for in high-stakes deployments.

Research angle

The interesting claim is that the win comes from systems engineering (invalidation, decay, bounded buffers) rather than a better model or training objective — a reminder that a large fraction of agent-memory quality is plumbing. Open: whether write-time invalidation generalizes outside the synthetic benchmark to noisy real market streams where "what is stale" is itself uncertain, and how the bounded buffer trades off against recall on rare-but-critical historical context.

→ Concept page: agent-memory