OPEN-DAG (Part I): Deterministic L1 Memory Invariants for Infinite-Horizon Multi-Agent Systems.

Abstract

Linear context accumulation is the dominant failure mode of contemporary multi-agent language-model systems. After N execution steps of mean residual length L, concatenating architectures occupy a resident working set of Θ(N·L), exhaust finite windows, and crash. We argue that this trajectory is an architectural choice, not a law of transformers. OPEN-DAG externalizes agent history into an immutable directed acyclic graph managed by a state manager on secondary storage, and pages a single skill vertex into a throttled L1 RAM working set. The resulting identity is Tdag(N)=600+H+400 , which is independent of execution depth. Under the evaluated configuration with H = 3,000 tokens, the active working set is 4,000 tokens. At N = 20 and S = 5 skills, the concatenating baseline occupies 24,000 tokens (83.3% saved). On a 15.50 GiB host the production L1 ceiling is 5% of DRAM (793.47 MiB). A 200 KiB factory demo throttle reached 91.9% occupancy (183.9 / 200.0 KiB) at depth 16 without a crash. Skills page from SSD into RAM under a circuit breaker; unused playbooks never enter the prompt. Pointer rollback completed in 3.80 ms versus 2,445 ms for prompt rewriting. These results demonstrate that linear context growth is not a necessary architectural property of long-horizon agent systems. Enterprise-scale memory hierarchy is deferred to a companion paper.