# SheetCompass: Hierarchical Relation Graphs for Agentic Spreadsheet Reasoning

Source: [arXiv](https://arxiv.org/abs/2608.14452v1)  
Feed7 permalink: https://feed7.dev/p/2608-14452v1-17xm2vb  
Published: 2026-08-14T16:39:16.000Z  
Trust: Needs Review (needs_review)

## Why Included

SheetCompass keeps spreadsheet structure as a hierarchical relation graph instead of flattening cells into text. Its agent memory retains task-relevant context across sheets, tables, and columns.

## Source Summary

SheetCompass models workbook relationships within and across worksheets using **hierarchical relation graphs**. A memory-driven agent retains task-relevant information while reasoning over cross-table associations, column dependencies, and spatial layouts.

## Practical Implication

For spreadsheet agents, preserve sheet boundaries and explicit structural links in the context representation. Treat the workbook as a connected data structure, then keep only task-relevant findings in agent memory as work progresses.

## Agent-Ready Context

SheetCompass models workbook relationships within and across worksheets using **hierarchical relation graphs**. A memory-driven agent retains task-relevant information while reasoning over cross-table associations, column dependencies, and spatial layouts.

For spreadsheet agents, preserve sheet boundaries and explicit structural links in the context representation. Treat the workbook as a connected data structure, then keep only task-relevant findings in agent memory as work progresses.

The supplied material describes the framework but gives **no benchmark results** or implementation details. It therefore supports the representation choice, not a measured claim that the approach outperforms existing spreadsheet agents.

## Connected Context

Feed7 judgment across 468 accumulated Signals:

SheetCompass specializes graph-shaped context for spreadsheets by preserving worksheet boundaries, cross-table links, column dependencies, and spatial layout, then using task-scoped memory to control what persists during reasoning. It reinforces structured retrieval over flattening a workbook into text, but does not establish superiority without benchmarks. Provenance, conflict handling, memory governance, and implementation cost remain outside the supplied framework.

- [AI on Your Lakehouse: Context Comes in Shapes, Not Queries — Zach Blumenfeld, Neo4j](https://feed7.dev/p/ai-on-your-lakehouse-context-comes-in-shapes-not-queries-zach-blumenfeld-0r4s9u5) — SheetCompass makes the candidate’s general graph-shaped-context recommendation concrete for workbooks, defining spreadsheet-specific relations that vector search or flat table querying may fail to preserve.
- [Citation Needed: Provenance for LLM-Built Knowledge Graphs — Daniel Chalef, Zep AI](https://feed7.dev/p/citation-needed-provenance-for-llm-built-knowledge-graphs-daniel-chalef-1iob5t8) — The provenance candidate adds a missing requirement to SheetCompass’s relation graph: structural links alone do not record how derived or merged facts trace back to workbook cells and sheets.
- [Lessons from Studying Every Memory System — Shlok Khemani, Independent](https://feed7.dev/p/lessons-from-studying-every-memory-system-shlok-khemani-independent-0m3gyxb) — SheetCompass uses task-relevant agent memory, while the memory-systems candidate warns that retained summaries need conflict detection, visibility, editing, and update policies that the framework description does not address.
- [WTF Is the Context Layer? The Missing Infrastructure for Production Agents — Prukalpa Sankar](https://feed7.dev/p/wtf-is-the-context-layer-the-missing-infrastructure-for-production-agent-0t47xqf) — SheetCompass emphasizes task-scoped memory inside a spreadsheet agent, whereas Atlan argues that durable business context should be shared and versioned across agents; this creates an architectural choice about where workbook-derived knowledge should persist.

## Context Map

- Layer: context
- Domains: data
- Topics: context-engineering, agent-memory, retrieval

## Uncertainty

- The supplied material describes the framework but gives **no benchmark results** or implementation details. It therefore supports the representation choice, not a measured claim that the approach outperforms existing spreadsheet agents.

## Agent Instruction

Use this item as source-backed context. Do not invent claims beyond the linked source. If this item conflicts with another source, call out the conflict.