MARS: Hierarchical Memory for LLM-Based Recommendation
A new framework called MARS (Memory-Augmented Agentic Recommender System) treats recommendation as a partially observable problem, using a structured belief state to abstract noisy user behavior into compact preference estimates. The system organizes memory into three tiers: event memory for raw signals, preference memory for mutable chunks with strength and evidence tracking, and profile memory for a coherent natural language narrative. It introduces a complete lifecycle of six operations—extraction, reinforcement, weakening, consolidation, forgetting, and resynthesis—to govern memory evolution. This approach contrasts with existing flat memory representations that conflate ephemeral signals with stable preferences. The work is published on arXiv under identifier 2605.14401.
Key facts
- MARS stands for Memory-Augmented Agentic Recommender System.
- It treats recommendation as a partially observable problem.
- The system maintains a structured belief state with three memory tiers: event, preference, and profile.
- Preference memory includes fine-grained mutable chunks with explicit strength and evidence tracking.
- Profile memory distills preferences into a natural language narrative.
- Six lifecycle operations govern memory: extraction, reinforcement, weakening, consolidation, forgetting, and resynthesis.
- The approach addresses limitations of flat memory representations in existing LLM agents.
- The paper is available on arXiv with ID 2605.14401.
Entities
Institutions
- arXiv