Research map / Large language models / Retrieval and knowledge
Knowing What is Missing: Efficient Conversational Memory via Explicit Evidence-Gap Tracking
Xingbo Du, Loka Li, Duzhen Zhang, Leonard Song, Le Song
Read the paperCheck your idea against 137,217 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 an efficient conversational memory system for LLM agents that explicitly tracks evidence gaps to guide multi-step retrieval over long dialogue histories, aiming to avoid context expansion and token cost.
Problem
Long-term conversational memory for LLM agents: full-history reasoning is token-expensive and noisy, while retrieval-only methods fail on multi-hop and temporal questions; existing multi-step retrievers grow the textual context over time.
Approach
Explicitly tracks which evidence is still missing for a question and uses those gaps to drive retrieval steps, instead of appending ever-growing raw textual history to the context. Details beyond the abstract are not available.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| wrap RAG backend with MemR3 evidence-gap controller | RAG · LoCoMo overall (GPT-4o-mini) | LLM-judge (%): 75.54 -> 81.55 | Improved |
| wrap RAG backend with MemR3 | RAG · LoCoMo overall (GPT-4.1-mini) | LLM-judge (%): 79.46 -> 86.75 | Improved |
| wrap Zep backend with MemR3 | Zep · LoCoMo overall | LLM-judge (%): 74.62 -> 76.26 (4o); 78.94 -> 80.88 (4.1) | Improved |
| add MemR3 to RAG | RAG · LoCoMo temporal (GPT-4o-mini) | LLM-judge (%): 65.11 -> 76.22 | Improved |
| add absorb-and-discard masking of raw snippets (ablation) | MemR3 (RAG) · LoCoMo overall | LLM-judge (%): 68.54 -> 81.55 | Improved |
| add state-driven refinement query delta q_k (ablation) | MemR3 (RAG) · LoCoMo overall | LLM-judge (%): 77.11 -> 81.55 | Improved |
| add reflect action (ablation) | MemR3 (RAG) · LoCoMo overall | LLM-judge (%): 76.65 -> 81.55 | Improved |
| use implicit multi-step retrievers (Self-RAG, RCR) instead of vanilla RAG | RAG · LongMemEval_s | accuracy | Hurt |
| use MemR3 instead of Full-Context | LLM backends · LongMemEval_s | accuracy | Mixed |
| use MemR3 instead of Full-Context | GPT-4o-mini · LongMemEval_s latency | avg latency (s): 11.04 -> 3.87 | Improved |
| replace model-estimated state with human-verified oracle state (ablation) | MemR3 · LoCoMo 152-question audit | human accuracy (%): 63.2 -> 84.2 | Improved |
| MemR3 versus Full-Context on LoCoMo overall | MemR3 (RAG) · LoCoMo overall (GPT-4.1-mini) | LLM-judge (%): 86.75 vs 89.00 | Hurt |
Experimental setup
Baselines: multi-step retrievers
Closest papers by meaning
- LongMemEval: Benchmarking Chat Assistants on Long-Term Interactive Memory (ICLR 2025 · similarity 0.90)
- SeCom: On Memory Construction and Retrieval for Personalized Conversational Agents (ICLR 2025 · similarity 0.89)
- From Isolated Conversations to Hierarchical Schemas: Dynamic Tree Memory Representation for LLMs (ICLR 2025 · similarity 0.89)
- PlugMem: A Task-Agnostic Plugin Memory Module for LLM Agents (ICML 2026 · similarity 0.89)
- Memory-T1: Reinforcement Learning for Temporal Reasoning in Multi-session Agents (ICLR 2026 · similarity 0.88)
- REMem: Reasoning with Episodic Memory in Language Agent (ICLR 2026 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
