Thousands of texts. Maybe years of them. Grey Mirror analyzes your entire exported history so you can compare the good stretches, the rough ones and what changed between them. Keep the dates and both speakers in the file. The report can only read what the export contains.
By Layne Weant
Published
Updated
A tool can let you upload a large export and still only reason over a thin slice of it. What matters is whether the product preserves the timeline and turns the thread into stable pattern measurements.
Repair failure, response-time drift, breadcrumbing, delayed accountability, and collapsing warmth usually become obvious only when the thread stays in sequence over time.
Grey Mirror can accept very large histories, and the tradeoff is time rather than thinner metric counts. Huge exports show a clear progress state while Grey Mirror works through the complete timeline.
Grey Mirror is built for exported conversations, zipped message archives, and long-running threads where the answer depends on what happened before and after the most dramatic moment. The upload workflow has no file size limit. Large files upload in chunks, and message coverage is checked before a report is published.
That matters because long-thread relationship analysis is not just an NLP token question. It is a sequence question. You want the thread to preserve who initiated, who withdrew, when the pacing changed, whether repair landed, and whether a dynamic improved or degraded over time.
Short excerpts are often enough for a red-flag triage question. They are weaker for questions like “who is carrying the repair,” “when did the reply pattern change,” “does this ex always reappear after silence,” or “is warmth still real or only occasional.” Those are long-context questions.
When someone asks for a large-context text analyzer, they are usually asking for pattern memory: the ability to compare earlier windows to later windows, detect cycles, and tell whether one reassuring message is part of real change or just another loop in a familiar pattern.
The product is built to accept very large histories, but it should not hide fidelity behind vague speed language. When a thread is huge, the user experience should say that processing can take longer because Grey Mirror is preserving the countable timeline, participant turns, response gaps, labels, and evidence windows.
Very large uploads should take longer rather than quietly thinning countable metric totals. The report should stay honest about processing time because count fidelity is more important than pretending a giant thread is instant.
That is a better promise than pretending large uploads are instant. The important distinction is this: metric counts, participant splits, response gaps, and taxonomy totals stay grounded in the parsed thread, while model-stage confidence is labeled clearly instead of silently weakening the report.
People looking for a large-context text analyzer often describe the same need in different ways: analyze long message threads, long text conversation analyzer, exported iMessage analyzer, WhatsApp chat analysis, years of texts analysis, or relationship thread analyzer.
Grey Mirror fits that need when the user wants full-thread relationship analysis, not just sentiment on a few lines. This page explains why thread depth matters before sending the reader into the tool.
Use the quick scan when you have a few pasted messages or screenshots and need fast triage. Use Grey Mirror when the question is pattern-heavy: long message thread analysis, repeated ex check-ins, repair decay, one-sided emotional labor, or a relationship that seems to have shifted gradually.
That separation keeps the product honest. The screenshot workflow answers a small-input question; Grey Mirror answers the large-context question.
You export the thread and upload it whole. That sounds obvious, and it is the step almost every other tool skips: screenshot analyzers ask you to choose which moments to show, which means the sample is selected by the person who already has a theory about what happened.
Reading the entire history removes that selection step. The ordinary weeks go in alongside the memorable arguments, which is the only way a timing or effort pattern can be separated from a bad fortnight. In the published sample the median thread ran 1,570 messages and the largest 163,636, and every one was read in a single pass rather than sampled.
Three things, and none of them survive being reduced to a screenshot. Recurrence, because a pattern is by definition something that happened more than once. Direction, because knowing whether reply times are lengthening requires the earlier baseline. And proportion, because "we argue constantly" is a claim with a number attached once the whole thread is counted.
The published figures make the point concretely: across 29 de-duplicated histories totalling 4,600,611 messages, about 97% of messages carried no detectable emotional signal and affection outnumbered explicit repair attempts by roughly ten to one. Neither number is knowable from an excerpt, and the ratio between them is the finding.
That is what this page exists for. The practical ceiling is not a message count, it is whether the export preserves sender and order. A 160,000-message archive with clean fields analyses cleanly; a short thread that lost its timestamps does not, and the report says so rather than guessing.
Very long histories take longer to process because the whole file is read rather than sampled, and the turning point detector runs a permutation test on every candidate so a long thread does not simply accumulate false positives. That test is why the benchmark reports 0% false positives on stationary threads rather than finding a turning point in everything it is given.
The upload workflow has no file size limit. Very large uploads should take longer rather than quietly thinning countable metric totals.
Every countable metric should stay grounded in the parsed message set. On huge exports, the report should show model-stage confidence instead of silently replacing the full thread with a thin sample.
People looking for a large-context text analyzer, long message thread analysis, exported chat analysis, or a tool that can read months or years of texts should land here first.
Read the layers
Visual guide updated
Keep the messages that actually exist in the export, without claiming access to missing material.
Sender mapping, chronology and aggregation must describe that same input throughout the report.
Each change or pattern should point to the relevant windows in the supplied history.
A high message count is valuable only when the report preserves context and evidence. Check the processed count and date range before trusting the summary.