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

E-mem: Multi-Agent Based Episodic Context Reconstruction for LLM Agent Memory

ICML 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 E-mem, a multi-agent memory framework where small-LM assistant agents keep raw episodic segments and reason locally over routed segments, while a master LLM aggregates their evidence, reporting SOTA F1 on LoCoMo and HotpotQA at lower token cost.

Problem

Preprocessing-based LLM agent memories (embeddings, graphs, summaries) compress sequential dependencies and lose context needed for multi-hop and temporal reasoning over long dialogues and document streams.

Approach

Instead of indexing compressed memory, each assistant agent keeps the uncompressed episodic segment and is activated by a multi-pathway router (global summary alignment, vector similarity, BM25 keyword match); activated assistants reason locally over raw text and return timestamped evidence to a master agent that aggregates it.

Key results

• On LoCoMo with GPT-4o-mini master, E-mem reaches 54.17 overall F1 vs 45.31 for GAM (the strongest baseline).
• On LoCoMo with Qwen2.5-14B master, E-mem reaches 57.04 overall F1 vs 50.41 for GAM.
• E-mem's normalized token cost is about 3.6k units vs 169k for long-context and 12.5k for GAM, with 11.25 s latency per query.

Surprising findings

• A simple direct-read pipeline without assistant agents collapses to 38.27 F1, far below the multi-agent design.
• On HotpotQA, standard RAG outperforms several complex memory baselines such as A-Mem, Mem0 and MemoryOS.

Limitations

• Higher latency (11.25 s per query) than retrieval-based baselines such as RAG, LightMem and Mem0.
• Cost comparison relies on an assumed 1:10 small-to-large token cost ratio.
• Long-context baseline with GPT-5.1 slightly beats E-mem on multi-hop and open-domain subsets.

Experimental setup

Datasets: LoCoMo HotpotQA

Benchmarks: LoCoMo HotpotQA AMA-bench

Metrics: F1 BLEU-1 normalized token cost latency

Baselines: Long-Context RAG A-Mem Mem0 MemoryOS LightMem GAM AriGraph HippoRAG

Builds on: GAM MemoryOS LightMem

Closest papers by meaning

• REMem: Reasoning with Episodic Memory in Language Agent (ICLR 2026 · similarity 0.90)
• Memory is Reconstructed, Not Retrieved: Graph Memory for LLM Agents (ICML 2026 · similarity 0.90)
• A-Mem: Agentic Memory for LLM Agents (NeurIPS 2025 · similarity 0.89)
• AMA-Bench: Evaluating Long-Horizon Memory for Agentic Applications (ICML 2026 · similarity 0.89)
• MemGen: Weaving Generative Latent Memory for Self-Evolving Agents (ICLR 2026 · similarity 0.89)
• EvoMemBench: Benchmarking Agent Memory from a Self-Evolving Perspective (NeurIPS 2026 · similarity 0.89)