CoVeMem: trainable collaborative vector memory for agents

简体中文

Core mechanism plus L2.1 oracle-free evaluation. Implements candidate-set retrieval, a vector-memory bank, and a soft-token read diagnostic.

Paper information

Field	Details
Paper	arXiv 2608.26895
Organization	Shenzhen Technology University, the first listed affiliation of the first author
First public release	2026-08-27 (arXiv v1)
Authors' code	No public author code was found when checked on 2026-08-29.
Adapter / method	covemem
Local reproduction code	src/auto_research/agent_research/latest_20260829.py

Original paper

Motivation and method

Text memory needs repeated LLM rewriting and loses full-catalog collaborative geometry. CoVeMem builds a bank from frozen LightGCN states, retrieves history relevant to the current candidate set, and projects it into soft tokens. Semantic alignment and masked listwise co-training teach the LLM to read memory.

flowchart LR
  G[LightGCN user and item states] --> B[Vector memory bank]
  C[Candidate set] --> R[Retrieve relevant history]
  B --> R
  R --> P[Projector soft tokens]
  P --> L[Masked listwise and pointwise readout]

Key figure from the paper

Figure 2: Overview of the paper's method and key components. From the original paper; copyright remains with the authors.

Core equations

\[ M_C=\operatorname{TopK}_{h\in H_u}\operatorname{sim}(e_h,\bar e_C),\qquad \mathcal L=\mathcal L_{align}+\lambda\mathcal L_{list}. \]

Paper results

CoVeMem matches or exceeds the strongest textual collaborative-memory method in 19 of 20 metric cells across four InstructRec benchmarks. Memory maintenance adds zero LLM calls after static profiling.

Local reproduction

Three ToolRoute-L2.1 seeds yield joint success 0.7708, plan F1 0.8004, and mean cost 4.5197.

See the metrics and batch index.

Reproduction boundaries

No LightGCN or LoRA checkpoint was trained. This verifies candidate-conditioned vector reads and control flow without text rewriting.