MARS: Memory-Aware Routing for Session-Level LLM Assistants
We introduce MARS, a memory-aware routing framework that uses session state as a shared control interface between memory construction and model selection.
ProofPaper ↗
Key points
- Long-term LLM assistants must repeatedly decide what prior context to expose and which model should answer as an interaction evolves.
- A Query and Topic Manager (QTM) estimates compact decision-relevant state that guides a fixed state-conditioned memory policy and an online router, enabling each turn to receive an appropriate memory view and a suitable model under a quality-memory-cost trade-off.
- We evaluate MARS on RealMem and MultiChallenge with 16 system configurations each, covering long-history memory-driven interaction and bounded sessions with structured context interference.
- Together, these results support a favorable quality-memory-cost trade-off for long-history, memory-dependent interaction.
Sources (1)
- [1]MARS: Memory-Aware Routing for Session-Level LLM AssistantsarXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 16, 03:04 PM
We introduce MARS, a memory-aware routing framework that uses session state as a shared control interface between memory construction and model selection.
Long-term LLM assistants must repeatedly decide what prior context to expose and which model should answer as an interaction evolves.
Extractive summary: sentences quoted from the sources.
Before this
- Oct 16, 2026Verbalising Model's Computation for Chain-of-Thought Faithfulness Detection
- Oct 16, 2026Liquid AI Releases d1-ultra: A Multimodal Decision Model That Scores 65 on the AA Decision Index
- Oct 16, 2026Scaling Indic Pretraining: 2.8 Trillion Tokens and an Open Data Generation Pipeline
- Oct 15, 2026Granite Switch 4.1: Chat, Decisions, and Speech from One Endpoint
- Oct 15, 2026NVIDIA Nemotron 3.5 Super VL Powers Specialized Agents with Multimodal Reasoning
- Oct 15, 2026Beyond Generation: Using Multimodal LLMs as Detectors of LLM-Generated Deepfake Facial Images