Title: RAMPART: Registry-based Agentic Memory with Priority-Aware Runtime Transformation

Abstract:

RAMPART introduces a novel compile-time memory architecture alongside a pure in-RAM block registry designed specifically for LLM-based agents. Within this framework, context assembly functions as a programmable runtime process, wherein content is compiled from a structured registry according to explicit policies governing ordering, inclusion, and eviction. This system utilizes five composable primitives—promote, gate, write, evict, and rollback—that operate on named, addressable blocks prior to compilation, incurring zero prompt-token costs. To establish a permissioned memory model, the system employs provenance tags and non-evictable authorship flags that enforce block-level ownership.

Experiments involving controlled probes with Qwen3-8B Q4 demonstrate that both compile-time block placement and the structural relationship between blocks and task queries significantly influence task success. Specifically, the performance cliff occurs at approximately the seventh block position when the task follows the registry structure, and at the twelfth position when it precedes it. Enhancing performance, grouping critical blocks with adjacent content and promoting the group as a single unit increases task success rates by tens of percentage points in scenarios where single-block placement is insufficient.

Cross-model replication across Qwen2.5-7B, Llama-3.1-8B, Mistral-7B-v0.3, and Qwen3-14B reveals that the content-priming effect manifests at consistent absolute positions across different model families, although the magnitude of this effect varies with model capability. Notably, block grouping boosts Mistral’s mean pass rate by approximately five times at the largest registry size. Furthermore, a smaller model utilizing this intervention can surpass a larger model lacking it in the mid-registry zone.

In terms of efficiency, relevance gating cuts prompt costs by 67.8% while recovering 83% of the success rate observed under promoted conditions. Additionally, schema eviction results in 0% invocations when the schema is absent, compared to 100% when present—a guarantee that policy-based approaches cannot inherently provide. Finally, shared-registry coordination minimizes inter-agent communication to a simple method call, achieving this with zero coordination token overhead.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC