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

M$^\star$: Every Task Deserves Its Own Memory Harness

NeurIPS 2026 · Multi-agent systems and games › LLM agents · approach: Agentic benchmarks and orchestration · 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.

Mstar discovers task-specific memory harnesses for LLM agents by evolving executable Python memory programs (Schema, Logic, Instruction) with a reflector coding agent and population-based search, outperforming nine memory baselines across LoCoMo, ALFWorld, HealthBench, and PRBench.

Problem

Fixed agent memory harnesses (storage, retrieval, and usage policies) transfer poorly across task types such as conversation, embodied planning, and expert reasoning. The paper asks how to automatically find the most suitable memory harness for a given task.

Approach

Represent a memory harness as an executable Python memory program with Schema, Logic, Instruction, and Toolkit components, then jointly optimize these via reflective code evolution with a coding-agent reflector, constraint-checked compile-fix repair, and softmax-sampled population search.

Key results

• On LoCoMo, Mstar reaches token F1 0.459 and LLM-judge 0.610, versus 0.373 and 0.540 for the best baseline (Mem0).
• On ALFWorld Unseen, Mstar achieves 0.881 success versus 0.857 for GEPA; on PRBench Finance it reaches 0.586 versus 0.449 for GEPA.
• Across five evolution seeds, 14 of 15 runs outperform the strongest baseline, with coefficient of variation at or below 10%.

Surprising findings

• Five of six cross-task transferred evolved programs perform worse than the universal seed baseline.
• Removing the code evolution (keeping only instruction evolution) causes the largest drop on LoCoMo (F1 0.459 to 0.256).

Limitations

• Evolution cost is substantial for rubric-graded benchmarks due to judge calls, and ALFWorld runs take about 100 hours of wall-clock time.
• Per-category gains are uneven; on LoCoMo, Mstar trails Mem0 and GEPA+Vector Search on temporal and causal questions.
• Only 20 evolution iterations and a small number of seed programs were explored.

Experimental setup

Datasets: LoCoMo ALFWorld HealthBench PRBench

Benchmarks: LoCoMo ALFWorld HealthBench PRBench

Metrics: token F1 LLM-judge score success rate rubric score

Baselines: No Memory Vector Search G-Memory Mem0 Trajectory Retrieval ReasoningBank Dynamic Cheatsheet GEPA GEPA + Vector Search

Builds on: GEPA AlphaEvolve-style evolutionary code search Vanilla RAG Mem0

Closest papers by meaning

• MemSkill: Learning and Evolving Memory Skills for Self-Evolving Agents (NeurIPS 2026 · similarity 0.91)
• EvoMemBench: Benchmarking Agent Memory from a Self-Evolving Perspective (NeurIPS 2026 · similarity 0.91)
• MemEvolve: Meta-Evolution of Agent Memory Systems (ICML 2026 · similarity 0.90)
• Evaluating Memory in LLM Agents via Incremental Multi-Turn Interactions (ICLR 2026 · similarity 0.90)
• MemGen: Weaving Generative Latent Memory for Self-Evolving Agents (ICLR 2026 · similarity 0.89)
• AdaMEM: Test-Time Adaptive Memory for Language Agents (ICML 2026 · similarity 0.89)