AuthorsYiming Du, Bingbing Wang, Yang He et al.
TL;DR
MemGuide uses intent-aligned retrieval plus missing-slot guided filtering to raise multi-session TOD success from 88% to 99% on MS-TOD while cutting 2.84 dialogue turns.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
Existing multi-session task-oriented dialogue relies on semantic similarity alone, ignoring task intent and slot continuity, which hurts coherence across sessions. Evaluations show retrieval-based methods only reach a 0.67 slot accuracy and 7.01 GPT-4 score, far below an oracle at 0.82 and 8.51.
This limitation breaks long-term LLM agents that must track evolving goals, forcing users to repeat slot values and inflating dialogue turns. As a result, multi-session TOD systems achieve only 88% task success, leaving a clear gap to intent-aware memory selection.
HOW IT WORKS
MemGuide centers on Intent-Aligned Retrieval, Missing-Slot Guided Filtering, a CoT Slot Reasoner, a LLaMA-8B filter, and an LLM Reader over QA memories to select useful history. MemGuide encodes intent keys with text-embedding-3-small and scores QA units by semantic similarity and marginal slot-completion gain.
You can think of MemGuide as a card catalog plus librarian: intent keys index shelves of QA memories, while the CoT Slot Reasoner and LLaMA-8B filter act like a librarian pulling only cards that fill missing details. Instead of dumping the whole archive into context, MemGuide surfaces just the few QA facts that close information gaps.
This design lets MemGuide exploit long-term memory beyond a plain context window by targeting missing slots explicitly. MemGuide can proactively confirm stored slot values, reduce redundant questions, and achieve minimal-turn, goal-consistent dialogue in multi-session TOD.
DIAGRAM
This diagram shows how MemGuide processes a user utterance through intent extraction, memory retrieval, missing-slot reasoning, filtering, and response generation across sessions.
DIAGRAM
This diagram shows how MS-TOD is constructed and how MemGuide is evaluated with confirmation-type responses and long-term memory.
PROCESS
Intent-Aligned Retrieval
MemGuide uses Intent-Aligned Retrieval to generate a current intent key with GPT-4o-mini and retrieve semantically closest intent descriptions and QA units from the memory bank.
Missing-Slot Guided Filtering
MemGuide runs Missing-Slot Guided Filtering by feeding the dialogue context and intent into the CoT Slot Reasoner to enumerate unresolved slots for the current task.
Re-ranking by Marginal Slot-Completion Gain
MemGuide applies the fine-tuned LLaMA-8B filter to score each QA pair by its probability of filling missing slots and combines this with semantic scores to select top K answers.
Response Generation
MemGuide passes the dialogue context, selected QA answers, and missing slot list to the LLM Reader, which generates proactive confirmation responses that minimize dialogue turns.
KEY CONTRIBUTIONS
MemGuide two-stage framework
MemGuide introduces Intent-Aligned Retrieval and Missing-Slot Guided Filtering with a CoT Slot Reasoner and LLaMA-8B filter, turning QA memory units into actionable context for minimal-turn task completion.
MS-TOD multi-session benchmark
MemGuide comes with MS-TOD, a benchmark of 2,861 dialogues, 18,530 utterances, 956 task goals, and 132 individuals, each averaging 21.67 sessions and 5.45 intents.
Intent-aware memory retrieval and filtering
MemGuide demonstrates that intent-aware QA memory and Missing-Slot Guided Filtering raise Success Rate to 0.99 and reduce dialogue turns by 2.84 compared to full-context prompting and traditional TOD systems.
RESULTS
GPT-4 score
7.14
+0.55 over ChatCite
Joint Goal Accuracy
0.70
+0.26 over AutoTOD
Dialogue Turn Efficiency
3.19
4.61 fewer turns than AutoTOD
Success Rate
0.99
+0.18 over ChatCite and +0.18 over AutoTOD
On the MS-TOD benchmark, which evaluates confirmation-type responses and long-term slot tracking, MemGuide with GPT-4o-mini achieves a GPT-4 score of 7.14, JGA of 0.70, DTE of 3.19, and Success Rate of 0.99. These numbers show that MemGuide converts intent-aware QA memory into concrete gains in both accuracy and interaction efficiency over AutoTOD and ChatCite.
BENCHMARK
On the MS-TOD benchmark, which evaluates confirmation-type responses and long-term slot tracking, MemGuide with GPT-4o-mini achieves a GPT-4 score of 7.14, JGA of 0.70, DTE of 3.19, and Success Rate of 0.99. These numbers show that MemGuide converts intent-aware QA memory into concrete gains in both accuracy and interaction efficiency over AutoTOD and ChatCite.
BENCHMARK
Success Rate on MS-TOD confirmation-type sessions.
BENCHMARK
Dialogue Turn Efficiency (DTE) on MS-TOD for GPT-4o-mini settings.
KEY INSIGHT
MemGuide shows that a smaller LLaMA-8B filter plus QA memory can raise JGA from 0.44 to 0.70 and Success Rate from 0.81 to 0.99. This happens even though the underlying generator, GPT-4o-mini, is unchanged.
This is surprising because many assume bigger generators or longer context alone solve multi-session TOD. MemGuide instead proves that structured QA memory and missing-slot guided filtering can deliver large gains without scaling the base LLM.
WHY IT MATTERS
MemGuide unlocks multi-session agents that remember user-specific slot values and proactively close information gaps with minimal extra turns. Developers can now build TOD systems that sustain 0.99 Success Rate across 20 plus sessions per user.
With MemGuide, builders can layer intent-aware QA memory and LLaMA-8B filtering on top of existing LLMs. This makes long-term, goal-consistent dialogue practical without exceeding context limits or retraining massive models.
Xingyu Lyu, Jianfeng He et al.
· 2026
ADAM combines Anchor extraction, Distribution estimation, Anchor selection, and Query generation to adaptively probe agent memory via an auxiliary generator and entropy based selection. On the EHRAgent benchmark with Llama2-7b-chat, ADAM reaches EQ=77 and ASR=1.00, compared to MEXTRA’s EQ=44 and ASR=0.89.
Okan Bursa
· 2026
Adaptive RAG Memory (ARM) augments a standard retriever–generator stack with a Dynamic Embedding Layer and Remembrance Engine that track usage statistics and apply selective remembrance and decay to embeddings. On a lightweight retrieval benchmark, ARM achieves NDCG@5 ≈ 0.9401 and Recall@5 = 1.000 with 22M parameters, matching larger baselines like gte-small while providing the best efficiency among ultra-efficient models.
Yijie Zhong, Yunfan Gao, Haofen Wang
· 2026
HingeMem combines Boundary Guided Long-Term Memory, Dialogue Boundary Extraction, Memory Construction, Query Adaptive Retrieval, Hyperedge Rerank, and Adaptive Stop to segment dialogues into element-indexed hyperedges and plan query-specific retrieval. On LOCOMO, HingeMem achieves 63.9 overall F1 and 75.1 LLM-as-a-Judge score, surpassing the best baseline Zep (56.9 F1) by 7.0 F1 without using category-specific QA formats.
Answers use this explainer on Memory Papers.
Checking…