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

MemEvolve: Meta-Evolution of Agent Memory Systems

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 MemEvolve, a bilevel meta-evolution framework that jointly evolves an agent's experience base and its modular memory architecture (encode/store/retrieve/manage), built on a unified codebase EvolveLab of twelve self-evolving memory systems.

Problem

Existing self-evolving agent memory systems use fixed, hand-designed architectures that do not adapt to task domains, leading to inconsistent gains across benchmarks.

Approach

Treat the memory architecture itself as an evolvable genotype: an inner loop accumulates experience under candidate memory systems, and an outer loop selects Pareto-top candidates (success, cost, latency) and uses LLM-driven diagnose-and-design to generate new modular memory implementations.

Key results

• MemEvolve on Flash-Searcher with GPT-5-mini reaches 74.71% on WebWalkerQA and 74.0% on xBench at pass@1, versus 71.18% and 69.0% for no-memory.
• Memory evolved on TaskCraft transfers to Kimi K2 and DeepSeek V3.2 backbones with gains of up to 17.06% on WebWalkerQA for Kimi K2.
• Among seven human-designed memory systems integrated with Flash-Searcher, MemEvolve yields consistent 3.54%-5.0% gains across three benchmarks while keeping cost near the no-memory baseline (GAIA $0.085 vs $0.086).

Surprising findings

• Many existing human-designed memory systems fail to give consistent gains; e.g., ExpeL underperforms on all three benchmarks and DILU degrades GAIA by 2.42%.
• Memory evolved on TaskCraft transfers across benchmarks, LLM backbones, and agent frameworks without manual adaptation.

Limitations

• Memory systems evolved on TaskCraft are unlikely to transfer to fundamentally different task families such as embodied action, per the authors.
• Evolved memory systems add per-task latency on some benchmarks (e.g., GAIA 693s vs 505s for no-memory), and the GAIA gains with the earlier variant are modest.
• Evaluation uses small subsets (e.g., 170 WebWalkerQA queries, 100 xBench tasks) and reports results on a limited set of frameworks and backbones.

Experimental setup

Datasets: GAIA TaskCraft WebWalkerQA xBench-DeepSearch

Benchmarks: GAIA WebWalkerQA xBench-DS TaskCraft

Metrics: pass@1 pass@2 pass@3 API cost per task delay per task number of steps

Baselines: Voyager ExpeL Generative DILU AWM Mobile-E Cheatsheet No-Memory AgentKB SmolAgent OWL Cognitive Kernel-Pro

Builds on: ExpeL Agent Workflow Memory Dynamic Cheatsheet AgentKB Flash-Searcher

Closest papers by meaning

• MemSkill: Learning and Evolving Memory Skills for Self-Evolving Agents (NeurIPS 2026 · similarity 0.94)
• EvoMemBench: Benchmarking Agent Memory from a Self-Evolving Perspective (NeurIPS 2026 · similarity 0.93)
• MemGen: Weaving Generative Latent Memory for Self-Evolving Agents (ICLR 2026 · similarity 0.93)
• MemRL: Self-Evolving Agents via Runtime Reinforcement Learning on Episodic Memory (NeurIPS 2026 · similarity 0.92)
• G-Memory: Tracing Hierarchical Memory for Multi-Agent Systems (NeurIPS 2025 · similarity 0.92)
• AdaMEM: Test-Time Adaptive Memory for Language Agents (ICML 2026 · similarity 0.92)