SuperLocalMemory 4.0 is presented as a local-first memory operating system for AI agents that unifies multi-channel retrieval through reciprocal-rank fusion, bi-temporal recall, scope isolation, role-based access, verified erasure, and hash-chained auditing. Its reliability spine controls the primary write path with generation-fenced admission, verifiable memory transactions, per-projection apply, verification, compensation and erasure owners, and hash-checkable completion manifests. In 11 fault-injection scenarios repeated 200 times, the system upheld 2,199 of 2,200 scoped component properties. The paper leads with a negative result: 10 mechanisms were implemented and reachable on live call paths but ineffective at their final connection, showing that implementation and reachability do not establish effectiveness without an independent oracle. It introduces a prior-distance assertion for Bayesian learners and a join-liveness assertion for schema-guarded paths, the latter identifying where missing guard data resides. A three-arm ablation that changed only the recall session-identifier namespace moved no posterior when the defect was present and moved every instantiated arm when it was absent; a negative control that wrote every ticket without engagement was inconclusive. The authors retract the previous version’s governed write-envelope overhead figure because the compared paths were not equivalent. In-place timing measured an 11.0 ms governed write, with the envelope accounting for 70.6%, while the generation fence cost 1.9 microseconds and the obligation ledger 42 microseconds. The paper characterizes the principal cost as durability rather than governance, and reports that version 2 retracts two version-1 claims and adds two controlled experiments.
