MemAnchor: A Parametric Memory Harness for Context Reconstruction

Language model agents need access to information beyond their active context window. We introduce MemAnchor, a parametric memory harness that trains a shared adapter to reconstruct source paragraphs from model-generated questions. At inference, the adapter is enabled only for recall; a frozen planner forms lookups and a frozen reader extracts answers for subsequent steps. On an exploratory set of 42 two-hop MuSiQue questions, MemAnchor answers 19 correctly (45.2% exact match), versus 17 (40.5%) for tuned two-step BM25 retrieval. Both configurations were selected on these questions; the small difference does not establish a reliable advantage. Disabling the adapter reduces exact match to 2.4%, but supplying gold paragraphs still leaves 16 questions answered incorrectly. Higher recall accuracy alone does not ensure better answers. Transfer is limited: retrieval performs better on HotpotQA, memory adds no measured benefit on LoCoMo, and sequential-writing exact match falls to 7.1% after ten rounds. These findings motivate studying memory writing together with its use, while leaving scalability and sustained long-horizon execution unresolved.