简体中文 · English

ReMem: Multimodal perception and evolving memory for recommender agents

L1 memory mechanism. Time-evolving memory with a fixed budget and a multi-memory GRPO objective are implemented. No real shopping pages or OCR model are connected.

Paper information

Field	Details
Paper	ReMem
Institution	The Hong Kong Polytechnic University (first author affiliation)
First public release	2026-09-29 (arXiv v1)
Original source code	Yes: Quhaoh233/ReMem
Local adapter/method	remem
Local reproduction code	src/auto_research/agent_research/latest_20260930.py

Original paper summary

Background and main changes

Recommender agents must extract information from heterogeneous product pages without adding unlimited user history to the context. ReMem uses OCR-driven multimodal item perception for compact semantic representations, then keeps preferences in dynamic memory with a fixed budget. Multi-Memory GRPO shares policy updates across memory views.

flowchart LR
  W[Product page or screenshot] --> P[OCR multimodal perception]
  H[History and interactions] --> M[Time-evolving memory]
  P --> A[Recommender agent]
  M --> A
  A --> G[Multi-Memory GRPO]
  G --> M

Key figure from the paper

Figure 1: The roles of multimodal “eyes” and a dynamic-memory “brain.” From the paper; copyright belongs to the authors.

Core mechanism

Local memory keeps recent observations and chooses historical blocks by decayed relevance/recency, with \(|M_t|\le B\). The objective computes group-normalized advantages separately for each memory view of the same query, then applies a shared clipped policy objective.

Offline and online results

The paper reports accuracy and token-efficiency gains for long-context recommender agents. Local checks only cover memory budget, temporal updates, the objective and gold isolation. See metrics.

Local reproduction and limits

This is not a complete web agent. No OCR, browser, author data or real episodes ran. Small fixed inputs check information flow, not recommendation capability gains.