Research map / Multi-agent systems and games / LLM agents

MemGuide: Intent-Driven Memory Selection for Goal-Oriented Multi-Session LLM Agents

AAAI 2026 · Multi-agent systems and games › LLM agents · approach: Self-evolving language agent learning

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes MemGuide, a two-stage intent-aligned memory retrieval plus missing-slot-guided filtering framework for multi-session task-oriented dialogue, and introduces the MS-TOD benchmark.

Problem

Semantic-similarity memory retrieval (RAG, long-context) ignores task intent and slot-level continuity, degrading coherence and efficiency in multi-session task-oriented dialogue.

Approach

Retrieve QA-formatted memory units by matching an LLM-extracted intent key to stored intent descriptions, then have a CoT slot reasoner enumerate missing slots and a fine-tuned LLaMA-8B filter re-rank QA units by their likelihood of filling those slots before proactive response generation.

Key results

• On MS-TOD, MemGuide raises success rate from 0.88 to 0.99 (GPT-4o-mini FCP vs MemGuide) and reduces DTE from 6.03 to 3.19 turns.
• With Mistral-7B as reader, MemGuide raises JGA from 0.73 to 0.80 and cuts DTE from 2.52 to 1.21.
• On SGD, MemGuide* reaches JGA 0.846 versus LDST 0.845; on MultiWOZ 2.2 it reaches JGA 0.879 versus TOATOD 0.638.

Surprising findings

• Retrieval-based memory outperforms full-context prompting on confirmation responses (dense retrieval 0.67 slot accuracy vs 0.61 for full context).
• MemGuide lowers BLEU and slot accuracy for several models (e.g., Mistral-7B BLEU 10.09 to 6.66 in one table) despite gains in GPT-4 score.

Limitations

• Dataset and code are promised only upon acceptance; evaluation relies on GPT-4 judging and simulated multi-session dialogues generated by GPT-4.
• Several results are inconsistent across tables (e.g., JGA and slot accuracy values differ between sections; BLEU drops).

Experimental setup

Datasets: Schema-Guided Dialogue (SGD) MS-TOD MultiWOZ 2.2

Benchmarks: MS-TOD SGD MultiWOZ 2.2

Metrics: GPT-4 score Joint Goal Accuracy Dialogue Turn Efficiency Success Rate human accuracy/informativeness/coherence slot accuracy BLEU ROUGE Recall@K

Baselines: Full-context prompting (FCP) Hybrid RAG AutoTOD ChatCite BERT-DST LDST BM25 retrieval text-embedding-3-small retrieval

Builds on: retrieval-augmented generation chain-of-thought prompting Schema-Guided Dialogue

Closest papers by meaning

• Evaluating Memory in LLM Agents via Incremental Multi-Turn Interactions (ICLR 2026 · similarity 0.88)
• Memory-T1: Reinforcement Learning for Temporal Reasoning in Multi-session Agents (ICLR 2026 · similarity 0.88)
• PlugMem: A Task-Agnostic Plugin Memory Module for LLM Agents (ICML 2026 · similarity 0.88)
• AMemGym: Interactive Memory Benchmarking for Assistants in Long-Horizon Conversations (ICLR 2026 · similarity 0.88)
• LongMemEval: Benchmarking Chat Assistants on Long-Term Interactive Memory (ICLR 2025 · similarity 0.87)
• AMA-Bench: Evaluating Long-Horizon Memory for Agentic Applications (ICML 2026 · similarity 0.87)