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

PlugMem: A Task-Agnostic Plugin Memory Module for LLM Agents

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 PlugMem, a task-agnostic plugin memory that converts raw episodic traces into a knowledge-centric graph of propositions and prescriptions, evaluated on LongMemEval, HotpotQA and WebArena with a PMI-based information-density metric.

Problem

Existing agent memory systems are either task-specific and non-transferable or task-agnostic but inefficient, suffering from low task relevance and context explosion when retrieving raw episodic memory.

Approach

Structure episodic memory into a graph whose nodes are knowledge units (propositional facts and procedural prescriptions) linked to high-level concept/intent nodes, enabling abstraction-to-specificity multi-hop retrieval and LLM-based condensation, with provenance back to episodes.

Key results

• On LongMemEval, PlugMem reaches 75.1% accuracy with about 363 memory tokens versus 63.6% for vanilla retrieval (3742 tokens) and 73.0% for LiCoMemory.
• On HotpotQA, PlugMem achieves 61.4 EM / 74.1 F1 with about 82 tokens, above HippoRAG2 (60.0 EM) and vanilla retrieval (51.7 EM).
• On WebArena Shopping online, PlugMem gets 52.6% SR versus 44.7% for A-Mem and 42.1% for AgentOccam, with 301 tokens on average.

Surprising findings

• Removing retrieval on HotpotQA drops EM to 20.0 and yields negative information density, with the reasoning module producing little context from randomly sampled items.

Limitations

• Relies on LLM-based extraction, abstraction and scoring (Qwen and GPT-4o), so memory construction cost and dependence on backbone quality are not fully characterized.
• Information density metric depends on estimating P_base and P_mem for the gold action, which requires gold optimal actions per decision instance.

Experimental setup

Datasets: LongMemEval HotpotQA WebArena

Benchmarks: LongMemEval HotpotQA WebArena

Metrics: accuracy exact match F1 success rate information density (PMI bits per token) average memory tokens

Baselines: Vanilla Retrieval A-Mem Zep LiCoMemory GraphRAG RAPTOR PropRAG HippoRAG2 AgentOccam AWM

Builds on: GraphRAG HippoRAG2 A-Mem RAPTOR

Closest papers by meaning

• A-Mem: Agentic Memory for LLM Agents (NeurIPS 2025 · similarity 0.92)
• Evaluating Memory in LLM Agents via Incremental Multi-Turn Interactions (ICLR 2026 · similarity 0.91)
• Harness the Memory: A Holistic Evaluation of Memory Substrates in Memory Agents (NeurIPS 2026 · similarity 0.91)
• EvoMemBench: Benchmarking Agent Memory from a Self-Evolving Perspective (NeurIPS 2026 · similarity 0.91)
• AMA-Bench: Evaluating Long-Horizon Memory for Agentic Applications (ICML 2026 · similarity 0.90)
• REMem: Reasoning with Episodic Memory in Language Agent (ICLR 2026 · similarity 0.89)