Why a General AI Chatbot Forgets Your Health History

You told a chatbot about your migraines last month. Today you ask about a new symptom, and it has no idea about the migraines — or it half-remembers them, without the dates. This guide explains, in plain language, why general-purpose AI chatbots handle your history the way they do, what newer memory features change, and what to look for if keeping a health history is the point.

Inside one conversation: the context window

A language model does not remember the way a person does. Each time it answers, it reads the text it has been given for that conversation and writes a reply. That text has a limit, called the context window.

Anthropic describes the context window as the model’s “working memory,” separate from the data it was trained on, and notes that as the amount of text grows, accuracy and recall degrade.^[1] In practice, that means the details you gave near the start of a very long conversation are the ones most at risk of being missed.

Between conversations: a fresh start, unless memory is on

A new conversation starts with a new context window. Unless the service copies something from your past into it, the model knows nothing about what you said before. That is why a general chatbot so often seems to “forget.”

That is changing. Since April 2025, ChatGPT has offered memory settings that let it draw on saved details and past conversations, which users can turn off and manage.^[3] In January 2026, OpenAI added a separate Health area designed to let people upload medical records and Apple Health data, with its own compartmentalized memories, which it says is not intended for diagnosis or treatment.^[4] So the honest answer today is not that chatbots always forget — it is that what they remember is decided by the system, in a form built for conversation rather than for a medical history.

Why memory isn’t the same as a record

• You may not see what was kept. Memory features keep the details the system selects. A record shows you every entry.
• Dates matter. “Has had headaches” is far less useful to a clinician than “headaches on 6 days in March, 2 in April.” Patterns live in the dates.
• Corrections should stick. In a record, you can fix a wrong entry and know it is fixed.
• It has to leave the chat. A history is only useful in an appointment if you can print it, export it or hand it over.

Your health history is already scattered: in 2024, 59% of people had more than one online medical record or patient portal, and only 7% used an app to bring them together.^[5] A chatbot’s memory adds one more partial copy unless it is designed as a record.

What to look for if you want an AI that keeps your history

• Can you see exactly what it has kept about you, as a list of dated entries?
• Can you correct or delete any single entry, and delete everything?
• Can you export it, or print a summary to take to an appointment?
• Does it tell you plainly who can access your information, and whether it is used to train AI models?
• Does it say clearly what it is not — that it does not diagnose and does not replace a clinician?

PocketHealth is built around those questions: what you report is saved to a dated timeline you can see, edit and delete, the assistant draws on it in later conversations, and Visit Prep turns it into a one-page summary. Our security page and consumer health data policy say who processes your information. Whatever tool you use, the guide to organizing medical records covers keeping a history of your own, and the guide to asking AI about symptoms covers where AI answers fall short.