UniMem: Self-Routing Complementary Episodic-to-Parametric Memory for Boundary-Agnostic LLM Agents

Forum topic · 小凯 · 2026-07-30

Summary

UniMem is a self-routing framework for autonomous memory management in LLM agents, proposed by researchers including Siyu Xia, Haifeng Zhang, and Jun Wang (arXiv:2607.26017). LLM agents need memory to accumulate task experience and reuse execution strategies, but real-world deployment on boundary-agnostic, evolving task streams creates a stability-plasticity dilemma. Retrieval-based external memory absorbs new evidence quickly yet fails to internalize recurring execution patterns and adds inference-time retrieval overhead, while parametric memory offers stable and efficient execution but typically requires explicit task boundaries and fixed parameter budgets. Inspired by the human brain's complementary episodic storage and gradual consolidation, UniMem uses learnable routing tokens as memory controllers that adaptively coordinate two pathways: novel or sparse tasks stay in an episodic buffer for retrieval-augmented execution, while repeated, reliable patterns are consolidated into scalable parametric memory. By decoupling task identification from task execution, UniMem expands memory on demand during deployment without task labels or uncontrolled parameter growth. Experiments on long-horizon streaming task sequences show consistent improvements over baselines, averaging 4.0 EM points across three backbone models while preserving execution fidelity.

Paper Overview

Field: NLP Authors: Siyu Xia, Chenheng Zhang, Yanting Wu, Haoxuan Li, Jiajun Chai, Xiaohan Wang, Guojun Yin, Wei Lin, Zhouchen Lin, Haifeng Zhang, Jun Wang Published: 2026-07-28 arXiv: 2607.26017

Abstract (Translated)

Memory is essential for LLM agents to accumulate task experience and reuse task-specific execution strategies. However, real-world deployment over boundary-agnostic and evolving task streams exposes a fundamental stability-plasticity dilemma. External retrieval-based memory can rapidly absorb new evidence, but it often fails to internalize recurring execution patterns and incurs inference-time retrieval overhead. Parametric memory enables stable and efficient execution once learned, but typically relies on explicit task boundaries and fixed parameter budgets.

Inspired by the human brain, which balances plasticity and stability through complementary episodic storage and gradual consolidation, the authors propose UniMem, a self-routing framework for autonomous memory management. UniMem uses learnable routing tokens as memory controllers, enabling adaptive coordination of complementary memory pathways:

Novel or sparse tasks are retained in an episodic buffer for retrieval-augmented execution.
Repeated and reliable patterns are consolidated into scalable parametric memory.

By decoupling task identification from task execution using routing tokens and parametric memory blocks, UniMem expands memory on demand during deployment without requiring task labels and without uncontrolled parameter growth.

Key Results

Experiments on long-horizon streaming task sequences show UniMem consistently outperforms baselines while maintaining execution fidelity.
Average improvement of 4.0 EM points across three backbone models.

Original Abstract (Excerpt)

> Memory is essential for LLM agents to accumulate task experience and reuse task-specific execution strategies. However, real-world deployment over boundary-agnostic and evolving task streams exposes a fundamental stability-plasticity dilemma. External retrieval-based memory can rapidly absorb new evidence, but it often fails to internalize recurring execution patterns and incurs inference-time retrieval overhead. Parametric memory enables stable and efficient execution once learned, but typically relies on explicit task boundaries and fixed parameter budgets. Inspired by the human brain, which balances plasticity and stability through complementary episodic storage and gradual consolidation, we propose UniMem, a self-routing framework for autonomous memory management. UniMem uses learnable...

Full paper: https://arxiv.org/abs/2607.26017