MemEvolve: Meta-Evolution of Agent Memory Systems — part 3/3

Cleaned-text code points [31214, 40185). Bibliography omitted here; source.md includes it.

MemEvolve Exhibits Cross-Task, Cross-Model, and Cross-Framework Generalization. Recall that the memory systems used on WebWalkerQA and xBench are directly inherited from those evolved on TaskCraft, without any task-specific meta-evolution. Nevertheless, these transferred memories yield consistent gains on more challenging benchmarks (WebWalkerQA+ : 58.82 → 61.18 % 58.82\rightarrow 61.18\% ; xBench+ : 69.0 → 74.0 % 69.0\rightarrow 74.0\% ), indicating that MemEvolve captures task-agnostic principles of memory design rather than overfitting to individual datasets. MemEvolve also demonstrates strong cross-LLM generalization. Although meta-evolution is conducted using GPT-5-Mini , memory systems evolved on TaskCraft+ transfer effectively to Kimi K2 and DeepSeek V3.2 without manual adaptation . Notably, Kimi K2 + improves by 17.06 % 17.06\% on WebWalkerQA and 10.0 % 10.0\% on TaskCraft. Finally, MemEvolve exhibits compelling cross-framework generalization. As shown in Figure 4 , directly transferring the memory system evolved on TaskCraft+ to heterogeneous agentic frameworks, including and , consistently improves performance despite substantial architectural differences. These results demonstrate that MemEvolve learns framework-agnostic memory abstractions that are readily pluggable across diverse agentic systems.

### 5.3 Self-Evolving Memory Comparison

We further compare the memory systems automatically evolved by MemEvolve against prevailing human-designed self-improving memory systems. In Table 3 , we integrate seven representative self-improving memory systems implemented in EvolveLab with Flash-Searcher, and comprehensively report performance, per-task cost/execution latency/execution steps. Results for MemEvolve are obtained using the system evolved on TaskCraft+ + GPT-5-Mini .

Existing Memory Systems Fail to Deliver Consistent Gains. Despite faithful re-implementations aligned with the original designs, many existing memory systems do not yield stable improvements. For example, DILU improves performance on xBench and WebWalkerQA, yet degrades GAIA by 2.42 % 2.42\% . Dynamic Cheatsheet achieves a 1.76 % 1.76\% gain on WebWalkerQA via skill condensation, but performs poorly on GAIA and xBench. More extreme cases are also observed: ExpeL underperforms on all three benchmarks. Upon closer inspection, this is unsurprising, as ExpeL was originally designed for relatively simple embodied or QA settings ( e.g. , ALFWorld, HotpotQA), and its prompts and mechanisms are ill-suited for long-horizon, long-context deep research. These results underscore the necessity of task-aware memory design.

MemEvolve Delivers Robust and Consistent Improvements. In contrast to prior approaches, MemEvolve yields stable and robust performance gains. Although the underlying memory system is evolved on TaskCraft, it consistently achieves improvements of 3.54 % ∼ 5.0 % 3.54\%\sim 5.0\% across all three evaluated benchmarks. Importantly, these gains are not achieved by substantially increasing the per-task cost. As shown in Table 3 , MemEvolve maintains API costs comparable to the No-Memory baseline across all benchmarks ( e.g. , GAIA: $0.085 vs. $0.086; xBench: $0.136 vs. $0.141), while its execution delay remains on a similar scale to other self-improving baselines ( e.g. , GAIA: 693.33 693.33 s vs. 584.88 584.88 s for AWM and 559.81 559.81 s for Cheatsheet; xBench: 773.06 773.06 s vs. 761.33 761.33 s for AWM and 818.07 818.07 s for Cheatsheet). Figure 5 further illustrates the cumulative success rate of different self-evolving memory systems as task execution progresses. Although performance exhibits higher variance in the early stages due to limited sample size, MemEvolve gradually stabilizes and converges to a consistently superior performance regime. This indicates that MemEvolve discovers principled and effective memory designs rather than relying on brittle, task-specific heuristics.

At first glance, such generalization may appear to conflict with our original motivation that memory systems cannot generalize across all domains and therefore require task-specific evolution. We argue this is not the case. Memory systems evolved on TaskCraft are unlikely to transfer effectively to fundamentally different task families ( e.g. , embodied action), where environments, action space and tool sets differ substantially. Nevertheless, MemEvolve enables the discovery of broadly applicable memory architectures within a shared task regime, while retaining the capacity for further task-specific adaptation when required.

### 5.4 Meta-Evolving Dynamics

Having established the substantial performance gains delivered by MemEvolve , we further examine how meta-evolution is executed in practice and which components are modified or enhanced during the evolutionary process. As illustrated in Figure 6 , MemEvolve starts from the predefined structure of AgentKB and iteratively evolves toward increasingly efficient memory architectures. Figures 10 and 9 highlights two high-performing memory systems discovered along this trajectory, denoted as Riva and Cerebra . Figure 8 presents a system evolved from the simplest few-shot example memory baseline, referred to as Lightweight .

Agents Spontaneously Evolve Efficient Memory Architectures. As illustrated in Figure 6 , the initial AgentKB memory system adopts a frozen design for both encoding and storage, lacking the capability to assimilate new experiences. Starting from this baseline, MemEvolve explores a spectrum of evolutionary directions. Some candidates are relatively aggressive (e.g., Ω 1 ( 1 ) \Omega^{(1)}_{1} , an Adaptive Decision System that decomposes a single agent trajectory into nine skill granularities), while others are more conservative (e.g., Ω 3 ( 1 ) \Omega^{(1)}_{3} , an Meta Memory System that stores trajectories at four levels and introduces an LLM-based meta-guardrail during retrieval to filter irrelevant information). The latter emerges as the winner in the first evolutionary round. The defining characteristic of this stage is agentic : both memory encoding and decoding increasingly rely on agent-driven decisions rather than predefined pipelines. The third evolution round introduces two further advances. Evolving from Ω 3 ( 2 ) \Omega^{(2)}_{3} Riva to Ω 1 ( 3 ) \Omega^{(3)}_{1} Cerebra , the memory system learns to distill not only textual insights but also reusable tools from past experience, while incorporating periodic maintenance of the memory database. Together, these enhancements provide faster evolutionary momentum for underlying agentic frameworks.

Evolved Memory Systems Are Effective in Practice. We further present concrete memory examples produced by the Lightweight system during real executions, as shown in Figure 7 . The results illustrate that Lightweight delivers memory content at varying levels of granularity, adaptively tailored to different task stages. During early planning, the memory provides high-level guidance, such as task decomposition strategies. As execution proceeds, it offers more fine-grained recommendations for tool-use, along with a form of working memory that highlights salient information from previous turns. Notably, Lightweight also exhibits predictive behavior by anticipating that target information may appear within image content on online travel websites, successfully guiding the agent to locate the evidence on trip.com. Together, these examples demonstrate the practical effectiveness of memory systems evolved by MemEvolve .

## 6 Conclusion

This work provides a unified implementation and design space for the rapidly growing field of self-evolving agent memory, together with a standardized codebase, termed EvolveLab , upon which we further build MemEvolve , a meta-evolutionary memory framework. Departing from the conventional paradigm of manually crafting a single self-improving memory architecture and expecting it to generalize across all domains, MemEvolve instead embraces adaptive, architecture-level evolution driven by empirical interaction feedback. Extensive experiments across diverse agentic benchmarks and backbone models demonstrate the effectiveness, robustness, and generalization of this approach. Moreover, analysis of the automatically evolved memory systems reveals several instructive design principles, including increased agentic involvement, hierarchical organization, and multi-level abstraction. We hope that MemEvolve serves as a step toward more automated, principled, and meta-evolutionary pathways for building continually improving agentic intelligence.

## 7 Contributions

Core Contributors • Guibin Zhang

• Haotian Ren

Contributors

• Chong Zhan

• Zhenhong Zhou

• Junhao Wang

• He Zhu

Corresponding Authors • Wangchunshu Zhou

• Shuicheng Yan

If you have any questions regarding the code, paper details, or other aspects of this work, you are very welcome to contact the authors at or via raising a Github issue.