Understanding human–AI collaboration through conversational data

By then, ChatGPT and Claude could both carry selected context across conversations. But memory was designed to improve the next response, not help users understand what had been retained or what patterns existed across their history. People could search old threads, but they could not see what they repeatedly returned to, what had changed, or what they had stopped asking about.

By May 2026, models could hold much more information at once, but became less reliable as that history grew. GraphWalks¹ is a benchmark that gives a model a large network of connected information and tests whether it can trace the right relationships through it. GPT-5.4 scored 93% with up to 128K tokens of context, but only 21.4% when the context grew to 256K–1M tokens²³.

My own archive contained nearly a thousand conversations and over 20,000 records across 28 months. The information existed, but it was scattered across hundreds of threads created for different purposes and in different states of mind. Neither the conversation archive nor the memory system could make sense of it as a whole.

Each part of the system was optimized for a different goal: models for capability, memory for the quality of the next response, and conversation interfaces for locating individual threads. No part was responsible for turning accumulated history into user understanding.

Mirror is deployed as a public prototype, with a downloadable offline copy that runs entirely in the browser.

Using Mirror with my own chat history confirmed patterns I already sensed, but gave me a richer view of where my attention had gone over time. The next question was whether the experience should end there or evolve into one people would return to.

Mirror currently analyzes everything in the browser. With much larger files or longer conversation histories, that processing could slow down or temporarily freeze the experience. I would need to stress-test those limits and explore other implementation approaches.

Because conversation data can be highly personal, a typical moderated session where someone moves through the experience and talks through their results would not be appropriate. I would separate testing the experience from testing the personal relevance of the report.

For moderated sessions, I would use realistic dummy data. This would allow participants to respond openly to the themes, structure, language, and overall experience without exposing their own conversations.

To understand whether Mirror produces useful insights from someone's actual history, I would have participants use it privately on their own device. After they had time to read the report, an optional rating flow or short questionnaire could ask what felt accurate, surprising, incomplete, or already known, as well as what they had expected the report to surface.

The larger question is how Mirror moves beyond the first report. Emerging features across conversational AI suggest that people may find value in both immediate usage insights and patterns drawn from their history. Feedback from the first report could help sharpen hypotheses for the next step: turning that history into ongoing, accumulated, and actionable insight.

Reflect provided an external signal in support of Mirror's initial hypothesis: that patterns across a person's conversation history can hold valuable insight beyond simply having that history remembered by the model. Anthropic's decision to invest in, build, and release the feature does not validate Mirror's specific design, but it suggests that this was a meaningful product opportunity beyond the project itself.

It also clarified a more important distinction: who gets to turn a person's usage history into value, and on whose terms. When that analysis is built into the model provider's product, the provider decides which patterns are surfaced, how they are interpreted, and how those insights feed back into the experience.

Mirror points to another possibility: people could take their conversation history outside the system that produced it and derive value from it through an independent layer of analysis. Foundation-model companies do not necessarily need to be the only ones defining which patterns become visible, what those patterns mean, or what users can do with them.