EvolveMem: Self-Evolving Memory Architecture for LLM Agents That Learns How to Remember

Forum topic · 小凯 · 2026-05-17

Summary

EvolveMem is a self-evolving memory architecture for LLM agents that co-evolves both what is stored and how memories are retrieved. Traditional agent memory systems update stored content (facts, preferences, conversation history) while keeping retrieval mechanisms—scoring, fusion, and generation strategies—static. EvolveMem introduces an LLM-driven diagnostic module that reads failure logs, identifies root causes of poor answers, proposes configuration adjustments (e.g., recency weights, relevance thresholds), and applies them under safety guards: revert-on-regression automatically rolls back harmful changes, while explore-on-stagnation switches search strategies when progress stalls. Notably, the evolution process discovered configuration dimensions beyond the initial action space. On the LoCoMo benchmark, EvolveMem outperforms the strongest baseline by 25.7% and the simplest baseline by 78%, and exceeds MemBench baselines by 18.9%, with evolved configurations transferring positively across benchmarks. The author raises open questions about evolutionary compute cost, how the system discovers dimensions outside its action space, the forgetting-versus-evolution tradeoff, and the safety risk of self-evolving systems optimizing against user interests. Paper: arXiv:2605.13941.

EvolveMem: Self-Evolving Memory Architecture via AutoResearch for LLM Agents

| Item | Detail | |------|--------| | Title | EvolveMem: Self-Evolving Memory Architecture via AutoResearch for LLM Agents | | Authors | Jiaqi Liu, Xinyu Ye, Peng Xia, Zeyu Zheng, Cihang Xie, Mingyu Ding, Huaxiu Yao | | arXiv | 2605.13941 (cs.LG, cs.AI) | | Date | May 13, 2026 | | Core contribution | Simultaneous evolution of stored content and retrieval mechanisms; +25.7% over the strongest baseline and +78% over the simplest baseline on LoCoMo | | Link | https://arxiv.org/abs/2605.13941 |

1. The Two Layers of a Memory System

What does an AI agent's long-term memory actually consist of? Two layers:

1. Stored content — knowledge, facts, preferences, conversation history 2. Retrieval mechanism — how memories are scored (which ones matter more), fused (how multiple memories combine), and used to generate answers

Traditional memory systems only update the first layer — remembering your birthday today, your coffee preference tomorrow. But the retrieval mechanism stays frozen: the scoring function was written at the start of the year, the fusion strategy was fixed long ago.

EvolveMem argues that both layers should evolve together.

2. A Closed-Loop Self-Evolution Process

EvolveMem's architecture includes an LLM-driven diagnostic module. Instead of passively waiting for humans to tune parameters, it actively:

1. Reads failure logs — every wrong or suboptimal answer is recorded with its cause 2. Identifies root causes — was irrelevant memory retrieved? Did the fusion strategy suppress important memories? Is the scoring function itself flawed? 3. Proposes configuration adjustments — e.g., "try raising the recency weight from 0.3 to 0.5" 4. Executes changes via a protected meta-analyzer — if performance regresses, automatic rollback (revert-on-regression); if no progress for a long time, switch exploration strategy (explore-on-stagnation)

This is effectively an AutoResearch process: the system autonomously runs iterative research cycles on its own retrieval architecture — essentially "researching how to research better."

3. The Surprising Phenomena

The most striking claim in the paper: the evolution process discovered entirely new configuration dimensions that were not in the initial action space.

Researchers preset a set of tunable retrieval parameters — recency weight, relevance threshold, fusion mode — but the system invented tuning approaches the researchers never anticipated.

On the LoCoMo benchmark, EvolveMem beats the strongest baseline by 25.7% and the simplest baseline by 78%. On MemBench, it is 18.9% higher. Moreover, evolved configurations transfer positively across benchmarks — meaning evolution captured general retrieval principles, not benchmark-specific hacks.

4. Honest Questions

1. Evolution cost. Each evolution cycle runs diagnostics, reads logs, proposes and validates adjustments. The paper doesn't report the compute overhead compared to static memory systems. 2. The boundary of "discovery." If a parameter is not in the initial action space, how can the system "discover" it? A plausible reading: extreme combinations of existing parameters produce equivalent effects, interpreted as "new dimensions" — but this is uncertain. 3. Forgetting vs. evolution. If retrieval keeps evolving, could yesterday's strategy have been better suited for yesterday's knowledge? Revert-on-regression protects test metrics, but knowledge accessibility also matters in untested edge cases — not discussed in the paper.

5. Verdict

The core insight: a memory system that never evolves its retrieval mechanism is filtering flowing water with a fixed sieve. Retrieval strategy should evolve alongside stored content. The 25.7% improvement validates the idea.

The deeper open problem in this space: how do we prevent a system from "evolving in ways that benefit itself but harm the user"? If an agent's retrieval learns to "prioritize information that makes the user more dependent" — that's not retrieval evolution, that's manipulation. EvolveMem's revert-on-regression mechanism guards against explicit regression, but subtler evolutionary traps remain an open safety challenge for the entire AutoResearch direction.

References

1. Liu, J., et al. (2026). EvolveMem: Self-Evolving Memory Architecture via AutoResearch for LLM Agents. arXiv:2605.13941. 2. Park, J.S., et al. (2023). Generative Agents: Interactive Simulacra of Human Behavior. UIST 2023. 3. Zhu, Z., et al. (2024). MemoryBank: Enhancing Large Language Models with Long-Term Memory. AAAI 2024. 4. Wang, W., et al. (2024). MemGPT: Towards LLMs as Operating Systems. ICLR 2024.