← Reddit

Claude in Excel using stale data from previous sessions, after introduction of chat history.

Reddit · EightFolding · May 26, 2026
A Claude in Excel user reported that the plugin has begun writing stale data from previous sessions instead of processing fresh CSV data, a malfunction that coincided with the introduction of chat history functionality. Despite Claude's attempts to resolve the issue through updated instruction sheets, the stale data problem persists across independent sessions and Excel restarts.

Detailed Analysis

Users of Claude's Excel integration are reporting a significant regression in functionality following the apparent introduction of persistent chat history to the plugin. The affected workflow involves a straightforward but high-value automation pattern: a user provides a CSV file, and Claude reads, parses, and writes that data into the active spreadsheet. What had functioned reliably for months across hundreds of executions began failing after chat history was introduced, with Claude consistently writing stale data from prior sessions into the workbook rather than the freshly provided input. The problem persists even when Excel is closed entirely, the workbook is reopened, and a new session is initiated — conditions that should, in theory, guarantee a clean environment.

The technical mechanism behind the failure is notable. Claude itself has been able to diagnose the issue when prompted, describing it as a situation where the `all_data` variable within its code execution environment retains values from a previous session, and where hardcoded arrays in `execute_office_jscalls` inherit that stale state. This suggests that the chat history feature, rather than simply providing conversational context to Claude, may also be inadvertently restoring or surfacing prior code execution state — or that Claude is drawing on prior session data in constructing new code blocks in a way that overrides fresh inputs. The distinction between conversational memory and execution environment state is a critical one, and conflating the two creates precisely this category of data integrity failure.

The broader implication is that the introduction of persistence features into agentic tool integrations carries non-obvious risks. When Claude is operating as an agent within a productivity application like Excel — reading files, executing code, and writing data — session state management becomes a safety and reliability concern, not merely a UX consideration. A user who trusts Claude to update a live workbook with fresh data and instead receives silently corrupted output from a prior session faces a potentially serious problem, particularly in financial, operational, or reporting contexts where spreadsheet accuracy is assumed. The user in this case notes that Claude's own self-correction attempts, including updating an instruction reference sheet within the workbook, have repeatedly failed to resolve the issue.

This incident reflects a wider challenge in deploying large language models as active agents in stateful environments. Unlike a simple Q&A interaction, agentic workflows require careful separation between contextual memory (what the model should know about the user's preferences and history) and execution state (what data is live and active in the current task). The Claude Excel plugin appears to be blurring this boundary in a way that its predecessor stateless version did not. As Anthropic and third-party integrators expand Claude's role in productivity software, robust state isolation protocols — ensuring that new task inputs cleanly supersede any cached or historical variables — will be essential to maintaining the kind of reliability that enterprise and power users depend on.

Read original article →