CoVeMem Replaces an Agent's Text Memory With Frozen LightGCN Vectors and Zero Maintenance LLM Calls

arXiv 2608.26895·medium signalXBlueskyLinkedInCopy link

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