Anthropic's memory export feature at claude.com/import-memory is not a dedicated data pipeline or structured API. It is a prompt.
The prompt instructs Claude to recall every stored memory verbatim, organized by date where available, covering user instructions, personal details, projects, tools, preferences, and corrections. It explicitly forbids summarizing, grouping, or omitting entries, and asks Claude to confirm completeness after output. This means the fidelity of your export depends entirely on the model's recall, not on a deterministic data retrieval system.
The implication is worth sitting with: a feature marketed around data portability is built on the same probabilistic text generation as everything else Claude does. Read the original to see the exact prompt text and decide for yourself what that means for memory reliability and user trust.
[READ ORIGINAL →]