Update Is a Set Operation, Retraction Is Not: A Multi-Scale Study of Memory Editing in Language Models

Large language models are increasingly used as long-term memory carriers, where two operations matter: updating a stale value and retracting it. We present a controlled, multi-scale study spanning four open-weight scales (Qwen2.5-0.5B-7B-Instruct), a second family (Yi-1.5-6B-Chat), a frontier API model (DeepSeek-V4.1-Flash), a real-data slice (LongMemEval), and a standard weight-edit audit, separating behavior, candidate set, residual, and reversibility. The results show a sharp asymmetry. (i) Storage-level writes behave as set replacement: KV-level write-time overwrite shows no rebound (0/20 at 0.5B-3B; 0/60 on a fresh pool), zero counterfactual leakage, no positive logit-lens layer, no old-value identity recoverable by a trained linear readout (0.5B/1.5B), and cross-family replication, while a standard ROME audit reads as replacement under the same instruments. (ii) Retraction text is carrier-dependent: appended retraction notices are ignored at every scale (0-13% flips), whereas false value assertions are accepted (49-60/60 locally, 60/60 at the frontier); a retraction compiled into the record is echoed rather than controlled (value-free refusals 4/0/1/13 per 60), and a frontier model that complies semantically still co-cites the retracted value in 100% of answers, with no value-free state to verify. On real data, prompting and labeling are inert while removing the superseded evidence raises accuracy by +31.8pp. (iii) Learned retraction is behavior-level and reversible across scales and families (18-20/20 behavior; 19/20 leakage; 20/20 revival). (iv) Constructive deletion via context reconstruction with authorization views achieves zero leakage and no linear residual. We conclude that update and retraction are different primitives: update rewrites the candidate set; retraction must control visibility, not add text. We release the instruments (P1-P5) as supplementary material.