CoVeMem replaces text memory with frozen LightGCN vectors and zero maintenance LLM calls.

Agentic recommenders keep user memory as narrative that further LLM calls rewrite one edit at a time, which makes exploiting full interaction history expensive and destroys graded collaborative similarity in the translation to sentences. CoVeMem makes the collaborative core a bank of frozen LightGCN user and item states, retrieved by the candidate set and injected as soft tokens beside a light textual profile. Across four instruction-grounded benchmarks it matches or beats the strongest text-memory agent on 19 of 20 metric cells with zero extra LLM calls for maintenance, against per-interaction calls for text memory. (arXiv 2608.26895)