Anthropic Gives Claude Users Full Memory Control, Locking Out Sensitive Topics by Default

2. Anthropic Gives Claude Users Full Memory Control, Locking Out Sensitive Topics by Default

On August 25, 2026, Anthropic announced a significant overhaul of how Claude handles persistent memory. Users can now view everything Claude remembers, organized topic by topic, and edit or delete any entry directly. Sensitive categories , health concerns, race, ethnicity, gender identity, religion, and political beliefs , are excluded from memory storage by default. A manual toggle lets users opt those subjects in, and anything mentioned before the toggle is flipped is never retroactively saved. Certain categories, including Social Security numbers, government IDs, criminal history, and immigration status, are permanently excluded regardless of settings. Memory now also persists across Claude's chat interface and Cowork, the company's desktop AI agent for multistep tasks.

The move is a direct competitive response to growing user distrust of opaque AI memory systems. OpenAI's Memory feature, which has drawn repeated criticism for storing information without clear user visibility, operates on a summarization model that users cannot easily audit or correct in real time. Anthropic's topic-by-topic architecture flips that model: memory is additive and immediate, updated as conversation happens rather than summarized after sessions end. That design choice matters for enterprise adoption, where IT and legal teams need demonstrable data controls. It also signals that Anthropic is betting trust architecture, not just model capability, as a product differentiator against both OpenAI and Google's Gemini, which has its own persistent context features but similarly limited granular controls.

The cross-surface memory sync between chat and Cowork is the detail worth watching most closely. Anthropic is quietly building a unified context layer that makes Cowork stickier the longer a user stays in the ecosystem. A manager's name, a team's metric definitions, a preferred report voice , these accumulate over months. That kind of long-running context is hard to replicate quickly, and it raises the switching cost for any competitor trying to pull users away from Claude's agent layer.

Source: Anthropic updates Claude's memory to enhance customization and protect sensitive topics