# E-mem: Multi-agent based Episodic Context Reconstruction for LLM Agent Memory

Source: [E-mem: Multi-agent based Episodic Context Reconstruction for LLM Agent Memory](<https://www.sota2.com/research/paper/697c3b5da8b01e76cc8da676>)

Collection index: [llms.txt](<https://www.sota2.com/research/llms.txt>)

Data updated: 2026-06-17T05:00:05.807Z

Authors: Kaixiang Wang, Yidan Lin, Jiong Lou, Zhaojiacheng Zhou, Bunyod Suvonov, Jie Li

Published: 2026-01-29T13:42:42.000Z

## Abstract

The evolution of Large Language Model (LLM) agents towards System~2 reasoning, characterized by deliberative, high-precision problem-solving, requires maintaining rigorous logical integrity over extended horizons. However, prevalent memory preprocessing paradigms suffer from destructive de-contextualization. By compressing complex sequential dependencies into pre-defined structures (e.g., embeddings or graphs), these methods sever the contextual integrity essential for deep reasoning. To address this, we propose E-mem, a framework shifting from Memory Preprocessing to Episodic Context Reconstruction. Inspired by biological engrams, E-mem employs a heterogeneous hierarchical architecture where multiple assistant agents maintain uncompressed memory contexts, while a central master agent orchestrates global planning. Unlike passive retrieval, our mechanism empowers assistants to locally reason within activated segments, extracting context-aware evidence before aggregation. Evaluations on the LoCoMo benchmark demonstrate that E-mem achieves over 54\\% F1, surpassing the state-of-the-art GAM by 7.75\\%, while reducing token cost by over 70\\%.

## Recorded benchmark results

A research result applies to its stated dataset, split, metric and evaluation configuration. Compare like-for-like settings, preserve the metric's unit and direction, and distinguish paper-reported results from independent reproduction. The best result recorded here is not a claim of exhaustive, current global coverage.

- [Multi-hop Question Answering on HotpotQA 400 documents](<https://www.sota2.com/research/sota/multi-hop-question-answering-on-hotpotqa-400-documents.md>): Multi-hop Question Answering; HotpotQA 400 documents; F1 Score: 61.46.

- [Multi-hop Question Answering on HotpotQA 800 documents](<https://www.sota2.com/research/sota/multi-hop-question-answering-on-hotpotqa-800-documents.md>): Multi-hop Question Answering; HotpotQA 800 documents; F1 Score: 55.46.

- [Multi-hop Question Answering on HotpotQA 1600 documents](<https://www.sota2.com/research/sota/multi-hop-question-answering-on-hotpotqa-1600-documents.md>): Multi-hop Question Answering; HotpotQA 1600 documents; F1: 55.76.

- [Complex High-Precision Reasoning on LoCoMo (test)](<https://www.sota2.com/research/sota/complex-high-precision-reasoning-on-locomo-test.md>): Complex High-Precision Reasoning; LoCoMo (test); F1 Score: 54.17.

- [Long-context reasoning and retrieval on LoCoMo (test)](<https://www.sota2.com/research/sota/long-context-reasoning-and-retrieval-on-locomo-test.md>): Long-context reasoning and retrieval; LoCoMo (test); Single-Hop F1: 61.14.

## Original sources and implementation

[Original paper](<https://arxiv.org/abs/2601.21714>)

[PDF](<https://arxiv.org/pdf/2601.21714.pdf>)