RPR-Memory: Decomposing Retrieval, Policy, and Reasoning in Persistent Memory Systems

Persistent-memory benchmarks for language-model-based agents are often reduced to one aggregate score. That scalar hides where a system failed: it may not retrieve the evidence, may expose the wrong memory state at read time, or may fail to compose the answer once evidence is available. We introduce RPR-Memory, a diagnostic protocol that reports these three layers separately: retrieval, read-time policy, and reasoning. We instantiate it with RPR-Bench, a benchmark harness that wraps heterogeneous memory architectures behind one six-primitive interface and drives them with three hard probes designed to fail at different pipeline stages. Across the frozen sweep, retrieval often succeeds before correctness: compositional-chain retrieval is near saturated while conditional accuracy remains variable under the fixed local composer. A post-freeze temporal ladder and human pilot localize the temporal-window case to model-side window-conjunction composition: on between-window rows humans move from 0.000 strict to 0.967 semantic EM, while the local model stays at 0.078 even with oracle-filtered candidates. The causal core is version resolution: adding write-time supersession to a controlled semantic-fact store lifts current-state update correctness from 0.000 to 0.680 but breaks historysensitive exact match (0.816 to 0.355); only query-conditioned history access at read time recovers both axes at once. That opposition makes read-time state exposure a separate, repairable layer of the pipeline, and an external Graphiti diagnostic corroborates the repair in a structurally distinct system. The contribution is methodological: a per-axis matrix that replaces the aggregate score, localizes which surface failed, and turns the diagnosis into a design prescription. Concurrent architectural work reaches a compatible conclusion independently.