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

REMem: Reasoning with Episodic Memory in Language Agent

ICLR 2026 · Multi-agent systems and games › LLM agents · approach: Self-evolving language agent learning · 0 citations (OpenAlex)

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 REMem, a two-phase episodic memory framework that builds a hybrid gist-and-fact graph with time-aware event nodes and uses an agentic ReAct-style retriever with curated tools, outperforming Mem0, Graphiti, and HippoRAG 2 on episodic recollection and reasoning benchmarks.

Problem

Language agents rely mainly on semantic memory and cannot effectively recollect and reason over episodic interaction histories with spatiotemporal context, since existing memory systems lack explicit event modeling and support for complex multi-event reasoning.

Approach

Index interactions as a hybrid memory graph of time-stamped gist nodes (event summaries) and time-scoped fact triples linked via context and synonymy edges, then let a ReAct-style agent iteratively call retrieval (semantic/BM25) and graph-exploration tools with temporal filters to reason over the graph.

Key results

• REMem-I achieves 42.4 F1 on LoCoMo vs 39.6 for NV-Embed-v2 and 39.0 for HippoRAG 2, with LLM-J 76.2.
• On Test of Time, REMem-I reaches 93.1% EM, the only method above 90%, versus 79.7% for Full-Context with GPT-4.1-mini.
• On Complex-TR, REMem-I obtains 89.6 LLM-J versus 80.4 for NV-Embed-v2 and 81.5 for HippoRAG 2.

Surprising findings

• Removing gists causes the largest drop on LoCoMo (LLM-J 76.2 to 48.9), while removing facts barely changes average LoCoMo results.
• The single-step REMem-S variant outperforms the iterative REMem-I on many LoCoMo single-session settings, as multi-step retrieval can add noise.

Limitations

• Errors dominated by selection/grounding mistakes and temporal or numerical reasoning errors on LoCoMo.
• Temporal window mismatch and multi-entity list incompleteness on Complex-TR.
• Indexing is offline batch; streaming memory construction is an open engineering challenge.

Experimental setup

Datasets: LoCoMo REALTALK Complex-TR Test of Time MuSiQue 2Wiki

Benchmarks: LoCoMo REALTALK Complex-TR Test of Time MuSiQue 2Wiki

Metrics: F1 BLEU-1 LLM-as-a-judge score exact match refusal precision/recall/F1

Baselines: Qwen3-Embedding-8B NV-Embed-v2 Mem0 Graphiti HippoRAG 2 TISER Full-Context

Builds on: HippoRAG 2 ReAct TISER Mem0

Closest papers by meaning

• Episodic Memories Generation and Evaluation Benchmark for Large Language Models (ICLR 2025 · similarity 0.92)
• MemGen: Weaving Generative Latent Memory for Self-Evolving Agents (ICLR 2026 · similarity 0.90)
• E-mem: Multi-Agent Based Episodic Context Reconstruction for LLM Agent Memory (ICML 2026 · similarity 0.90)
• Memory-T1: Reinforcement Learning for Temporal Reasoning in Multi-session Agents (ICLR 2026 · similarity 0.90)
• WorldMemArena: Evaluating Multimodal Agent Memory Through Action–World Interaction (NeurIPS 2026 · similarity 0.90)
• Human-inspired Episodic Memory for Infinite Context LLMs (ICLR 2025 · similarity 0.90)