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

Embodied Agents Meet Personalization: Investigating Challenges and Solutions Through the Lens of Memory Utilization

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.

Introduces MEMENTO, a two-stage evaluation of LLM embodied agents' episodic-memory utilization for personalized object rearrangement, and proposes a hierarchical knowledge-graph user-profile memory that improves performance.

Problem

LLM-powered embodied agents are evaluated on conventional rearrangement with explicit instructions, but personalized assistance requires recalling user-specific object semantics and behavioral routines from past interactions; it is unclear how well agents use such episodic memory.

Approach

A two-stage (acquisition then utilization) evaluation that shares scene and goal while varying instructions so that the underspecified utilization instruction can only be solved by recalling memory; plus a three-level hierarchical knowledge-graph user-profile memory with temporal edges for routines, added alongside episodic memory.

Key results

• GPT-4o success rate drops from 95.0 (acquisition) to 63.9 on joint-memory utilization tasks (-30.5 points).
• Llama-3.1-70b joint-memory success rate falls to 8.3% from 90.0% in acquisition.
• Adding user-profile memory yields substantial gains on single and joint-memory tasks for GPT-4o and Qwen-2.5-72b, with largest gains on user-pattern tasks.

Surprising findings

• Summarizing episodic memory gives little benefit for large models and hurts smaller models, showing trajectories provide in-context learning value beyond personalized knowledge.
• Agents with two object-semantic memories in joint tasks degrade notably (GPT-4o -20.8% SR), despite handling single object semantics well.

Limitations

• Experiments use a simulator with oracle perception and motor skills, excluding visual perception.
• No conversational component; instructions are templated and synthetic from GPT-4o.
• Evaluated mostly on a small set of LLMs with only two models for the user-profile memory experiment.

Experimental setup

Datasets: PartNR test set OVMM object models Google Scanned Objects

Benchmarks: MEMENTO

Metrics: Percent Complete Success Rate Sim Steps Planning Cycles delta PC delta SR

Baselines: full episodic memory summarized memory instruction-only

Builds on: ReAct PartNR Habitat 3.0 all-mpnet-base-v2

Closest papers by meaning

• MemGen: Weaving Generative Latent Memory for Self-Evolving Agents (ICLR 2026 · similarity 0.90)
• Personalize-then-Store: Benchmarking and Learning Personalized Memory for Long-Horizon Agents (NeurIPS 2026 · similarity 0.90)
• G-Memory: Tracing Hierarchical Memory for Multi-Agent Systems (NeurIPS 2025 · similarity 0.89)
• MemEvolve: Meta-Evolution of Agent Memory Systems (ICML 2026 · similarity 0.89)
• MemSkill: Learning and Evolving Memory Skills for Self-Evolving Agents (NeurIPS 2026 · similarity 0.89)
• A-Mem: Agentic Memory for LLM Agents (NeurIPS 2025 · similarity 0.89)