Control-Plane Placement Shapes Forgetting: An Architectural Study of Agent Memory Across Thirteen System Configurations

The position of a large language model inside an agent memory pipeline determines which forgetting failures the system can recover, according to a study posted to arXiv on June 14. Researchers compared thirteen configurations and found that no single placement covers all failure modes [1]. The paper, titled “Control-Plane Placement Shapes Forgetting,” examines where an LLM sits between the recall plane—which retrieves stored facts—and the control plane, which mutates them through operations such as supersede, release, and purge [1][2]. The recall plane has been extensively benchmarked, but the control plane remains largely untested [2]. The authors evaluated the thirteen configurations against a 385-case adversarial surface [1]. Deterministic primitives handled lexical and temporal categories but stumbled on canonicalization, posting a 5% failure rate on identifier-obfuscation and a 0% failure rate on cross-lingual tasks [2]. Placing the LLM at inscribe time recovered canonicalization completely—100%—yet offered no help on intent-aware deletion, where it recorded a 0% recovery rate on prefix-collision and compound-fact cases [2]. A mutation-time hook reversed that pattern. It recovered intent-aware deletion at a rate of 78–85% and lifted overall recovery to 91.7–93.2% [2]. The cost was modest: $0.17 per full 385-case run, with mutation latency of 2.3 seconds per case compared with 64–191 milliseconds for the deterministic recall path, which remained unchanged [2]. The evaluation harness, called ForgetEval, combines a 1,000-case templated suite with the 385-case adversarial layer—132 hand-crafted cases and 253 LLM-drafted, oracle-validated cases—scored by deterministic substring match [1][2]. A six-method Adapter Protocol with honest N/A scoring allows heterogeneous memory stores to enter in roughly 130 lines of code [2]. Ten annotators reviewed the admission criteria, producing a Fleiss’ kappa of 0.958, and a 77-case subset authored by four blind external contributors replicated the canonicalization asymmetry while amplifying the joint-placement lift by 27.8 percentage points [2]. The study, released under an MIT license, argues that production failures are predominantly forgetting failures rather than recall failures, even though existing benchmarks measure only recall [1][2]. arXiv, the open-access repository where the paper appeared, hosts e-prints across mathematics, physics, computer science, and related fields and has grown to a submission rate of about 24,000 articles per month as of late 2024 [6].