T1. Fidelity: "delete the source, keep the knowledge"

Anamnesis restores near-oracle answer quality at roughly an 18x token discount.

Each task asks for a specific, non-derivable fact about the system, chosen so that a model cannot guess it from reasoning alone (validated: a blinded model scores near zero). Single-message API calls isolate exactly what the model has in context.

• Arm A (Oracle): the source document is loaded into context. 98.3% fidelity, 7,820 tokens.
• Arm B (Blinded): no memory. 7.5% fidelity, 143 tokens.
• Arm C (Anamnesis): the relevant memory is injected. 94.2% fidelity, 431 tokens.

Memory closes almost the entire quality gap of the full-document baseline while carrying a fraction of the context. Scored blind against a fixed rubric, mean of 8 tasks over 5 runs each.