arXivPaperNeeds Review

SheetCompass: Hierarchical Relation Graphs for Agentic Spreadsheet Reasoning

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.

arXiv · Aug 14, 2026

Open Source Open Markdown Open JSON

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

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, Neo4jSheetCompass 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 AIThe 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, IndependentSheetCompass 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 SankarSheetCompass 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

contextdata#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.