Claude's 'Use account memory' toggle is on by default for new projects and enabling it is retroactive, contradicting Help Center docs

Testing of Claude's updated memory features reveals that the "Use account memory" toggle is enabled by default for new projects and was found enabled on a project created in 2024 that had never been modified. This setting allows regular, non-project chats to read memory files stored within specific projects, a behavior that contradicts current Help Center documentation stating that project memory is separate from general chats.

Testing Methodology and Findings

The user employed canary codes (random strings like CANARY-XXXXXXXX) placed in different contexts to verify data isolation:

• Default State: The toggle is on for new projects and was found enabled on a project created in 2024 that had never been modified.
• Data Leakage: When the toggle is on, regular chats can access project memory, and projects can access general memory. Switching the toggle on is retroactive; previously isolated memory became immediately visible to external chats.
• Isolation: Turning the toggle off successfully blocks both directions of memory access. Chat history remains separate, but distilled memory (decisions, contacts, status) crosses over.
• Model Awareness: Claude does not inherently know it has access to this data; it failed to find information until explicitly instructed to look in the project's memory.

Documentation Discrepancy

The Help Center articles (updated recently) claim that project memory is isolated from non-project chats. The observed behavior—where dozens of projects had auto-generated memory files readable from regular chats—suggests a gap between the documented privacy model and the actual default configuration. The user notes that shared memory is useful but flags the default-on setting as a potential privacy concern for users storing sensitive client work or NDAs in projects.