CoVeMem is a new memory design for agentic recommender systems, which use an LLM to make recommendations based on a persistent user memory. Instead of storing and updating a textual narrative, which is expensive and loses collaborative structure, CoVeMem uses vector states from a frozen LightGCN model as a memory bank. At each decision, the candidate set retrieves relevant historical states, which are projected into soft tokens for the LLM, alongside a static textual profile.
Training uses and masked listwise co-training to teach the model to rank through these collaborative vectors. Across four benchmarks, CoVeMem matches or exceeds the best text-memory agent on 19 of 20 metric cells, while eliminating per-interaction LLM calls for memory maintenance. This makes the full interaction history available as training data for the memory itself.