RevBound: Revocation Cuts for Transformed Agent Memory
Desk rejected at ICLR 2027, so left out of the paper lists and the map.
Abstract
An agent reads an authorized memory fragment, compresses it into a plan, and caches that plan for a peer. Access is then revoked. The source disappears from retrieval, yet its derivative can still enter the next model call. RevBound treats this lifecycle as a path-coverage problem over execution provenance: a revocation cut must intersect every denied-source path after its last mutable transformation. The Object-Alignment Principle identifies when revocation evaluations protect the same output under the same relational source closure, policy version, and decision boundary. RevBound realizes such a cut at model-call admission with typed user–agent and agent–resource lineage, a linearizable policy fence, transitive parent closure, and an AllowedBundle-only invocation path. In 1,160 matched fixed-parent attacks spanning re-summarization, paraphrase, merging, and relay, retrieval-time filtering exposes 600 forbidden derivatives (51.7%); two independently implemented final-boundary serializers expose 56 (4.8%) and 36 (3.1%), while RevBound exposes none (one-sided 95% upper bound \(\leq0.26%\)). RevBound also records zero violations over 156,000 churned bundle reads and 1.8M fragment checks at 7.6 ms p95. On SWE-Coop it resolves 28.4\(\pm\)1.1% of issues, compared with 23.6\(\pm\)1.0% for shared vector memory, and a second model/tool stack reproduces the boundary-class gap, with trace-audited leaks falling from 55/300 to 1/300. Revocation for transforming agent memory must therefore be enforced on the model-admitted object, where complete lineage and non-bypassable routing turn source denial into descendant exclusion.