HasMem Improves Long-Term Agent Memory via Adaptive Softening

HasMem uses a controller to adjust memory widths and a Writer module to re-encode resized entries for frozen LLMs. In reconstruction probes, the system achieved a 95.3 lexical F1 score, outperforming rule-based re-encoding by up to 23.6 percentage points in exact-match accuracy.