Detailed Analysis
A user on a Reddit forum devoted to Claude has surfaced a frustrating and consequential inconsistency in how Anthropic's memory feature handles sensitive personal data. The poster, who describes living with chronic illness and associated brain fog, had relied on Claude's memory to retain health details between specialist appointments, using the assistant to help track symptoms and produce structured summaries for doctors. According to the post, this workflow functioned reliably until a recent update—seemingly coinciding with a version the poster refers to as "Fable/Opus 5"—after which Claude began refusing to add new health information to memory, either through direct requests or organically during conversation. Notably, the user reports that a "health" category already exists within their memory settings, populated with some information from recent chats, yet the model still declines to append anything further to that same section, creating a confusing and self-contradictory experience: partial retention without editability.
The underlying tension here is between safety-motivated guardrails and genuine user autonomy over personal data. Anthropic, like other AI labs, has built in restrictions around storing and processing sensitive categories of information—health, biometric, or otherwise legally protected data—largely to reduce liability and avoid the appearance of the model acting as a de facto medical record-keeper or advice-giver. But as this case illustrates, blanket restrictions applied inconsistently can actively harm the very users the policy is meant to protect. For someone with cognitive impairment from chronic illness, the ability to offload memory work to an AI assistant is not a convenience but a functional necessity, and a sudden, unexplained policy shift removes a coping mechanism without warning or clear recourse. The user's frustration is compounded by the fact that they are a paying subscriber at a premium tier (£250/month, suggesting a Max or enterprise-level plan), which sharpens the expectation that they should have control over their own stored data rather than having a third party unilaterally curate what can and cannot be remembered.
This incident also reflects a broader pattern in consumer AI: the friction between centralized safety policies and the long tail of legitimate, high-stakes personal use cases that don't fit neatly into "general assistant" design assumptions. Memory features across AI products (ChatGPT's memory, Claude's memory, Gemini's personalization) are relatively new and still being tuned, often through opaque backend changes that ship without user-facing changelogs—hence the poster's confusion about "when this new change has come in." Health data specifically sits at the intersection of HIPAA-adjacent sensitivities, litigation risk, and genuine therapeutic value, making it a prime candidate for overcautious filtering that satisfies legal teams but frustrates the disabled and chronically ill users who most need durable, low-friction memory tools. The workaround culture visible in the post itself—users turning to Obsidian vaults, Claude Projects, and file uploads to route around memory restrictions—shows that determined users can reconstruct the functionality manually, but this defeats the purpose of an integrated memory feature and disproportionately burdens exactly the people (those with cognitive fog, fatigue, or disability) least able to manage complex workarounds.
More broadly, this case is a small but telling data point in the ongoing debate about who should control an AI's memory of you: the user, the platform, or some negotiated combination of both. As Anthropic and competitors push memory as a flagship differentiator for long-term assistant relationships, incidents like this will likely accelerate calls for more granular user controls—letting individuals explicitly consent to storing categories like health information, rather than having a system silently decide what's too sensitive to retain. Until such controls exist, sudden and unexplained changes to memory behavior will continue to erode trust among the power users, including disabled and chronically ill communities, who have come to depend on these assistants as accessibility tools rather than mere novelties.
Read original article →