Captured: Claude — Memory Layer Discussion

You: I'm thinking about authored memory vs extracted memory. Where does extracted memory fail?

Claude: Three places, in order of severity:

1. Misextraction — the model picks up a passing comment as a "preference" and re-inserts it forever. Hard to undo because the user doesn't know it's there.

2. Walled gardens — extracted memory lives in the vendor's DB. Switch vendors → memory doesn't follow. Lock-in by accident.

3. Audit trail — extracted memory has no edit log the user can reason about. Authored memory has commits, diffs, version history.

You: So when does extracted win?

Claude: Casual personal use. "Remember that I'm vegetarian" doesn't need a versioning system. Authored memory wins for high-stakes, multi-AI, or long-lived knowledge.