arxiv: 2604.10923 · v1 · submitted 2026-04-13 · 💻 cs.CL · cs.AI

Mem²Evolve: Towards Self-Evolving Agents via Co-Evolutionary Capability Expansion and Experience Distillation

Zihao Cheng , Zeming Liu , Yingyu Shan , Xinyi Wang , Xiangrong Zhu , Yunpu Ma , Hongru Wang , Yuhang Guo

show 2 more authors

Wei Lin Yunhong Wang

This is my paper

Pith reviewed 2026-05-10 16:40 UTC · model grok-4.3

classification 💻 cs.CL cs.AI

keywords self-evolving agentsexperience memoryasset memorycapability expansionexperience distillationLLM agentsco-evolutionary agents

0 comments p. Extension

The pith

Agents that co-evolve experience and new assets outperform those using either process in isolation.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper proposes that self-evolving language model agents reach greater and more stable capability gains when they use accumulated experience to guide the creation of new tools and expert agents, while those new assets in turn supply fresh experience for further refinement. Standard approaches either accumulate experience within a fixed set of tools or create new assets without experiential guidance, both of which limit growth. Mem²Evolve combines an Experience Memory and an Asset Memory to create a mutual reinforcement loop. Experiments on eight benchmarks across six task types show consistent gains, suggesting the co-evolutionary approach overcomes the bounds of static toolsets and unguided asset generation.

Core claim

Mem²Evolve integrates Experience Memory, which stores task outcomes and strategies, with Asset Memory, which dynamically generates and retains new tools and specialized agents. By using experience to inform asset creation and new assets to generate novel experiences, the framework achieves co-evolutionary expansion of agent capabilities. This results in higher performance across diverse tasks compared to baselines that evolve through experience alone or asset creation alone.

What carries the argument

The dual-memory architecture consisting of Experience Memory for distilling past interactions and Asset Memory for expanding the set of usable tools and agents, linked through a co-evolutionary process.

If this is right

Agents can expand their capabilities beyond any manually predefined static toolset.
The evolution process becomes more stable by avoiding unguided asset creation.
Experience from new assets enriches the memory for future guidance.
Overall performance improves by double digits over isolated evolution methods on multiple benchmarks.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Similar mutual reinforcement between memory types could be applied to other self-improving systems like code generation or planning agents.
Over very long horizons the co-evolution might encounter diminishing returns that require additional mechanisms to sustain.
Deploying such agents in open-ended environments could test whether the gains transfer beyond the controlled benchmarks used here.

Load-bearing premise

The two evolutionary processes of experience accumulation and asset creation are intrinsically interdependent in a way that produces stable, unbounded growth rather than interference or diminishing returns.

What would settle it

A controlled test that runs the Mem²Evolve process over many more iterations than reported and checks whether performance gains continue, level off, or reverse.

Figures

Figures reproduced from arXiv: 2604.10923 by Hongru Wang, Wei Lin, Xiangrong Zhu, Xinyi Wang, Yingyu Shan, Yuhang Guo, Yunhong Wang, Yunpu Ma, Zeming Liu, Zihao Cheng.

**Figure 1.** Figure 1: Paradigms of Self-Evolving Agents: (a) Experience-centric evolution, (b) Capability-centric evolution, and (c) Our co-evolutionary framework that jointly expands capabilities and distills experience. capabilities (Gao et al., 2025; Fang et al., 2025; Wang et al., 2025a). However, current frameworks predominantly treat these evolutionary processes in isolation (Cemri et al., 2025). As illustrated in Figu… view at source ↗

**Figure 2.** Figure 2: Overview of Mem2Evolve, a self-evolving agent framework built on a Dual-Memory mechanism. The evolution proceeds in two phases. During Forward Inference, the agent recruits tools and expert agents from Asset Memory to execute the current task. When the task exceeds its current capability boundary, Experience Memory is leveraged to guide the stable creation of new assets on demand. During Backward Evolution… view at source ↗

**Figure 4.** Figure 4: Cross-task self-evolving performance. When initialized with heterogeneous memory from GAIA, Mem2Evolve consistently outperforms the setting without initial memory and achieves performance comparable to single-task initialization. To evaluate the generalization capability of Mem2Evolve in a cross-task setting, we initialize the agent with heterogeneous memory accumulated from GAIA and evaluate its performa… view at source ↗

**Figure 3.** Figure 3: Single-task self-evolving performance. The results show that initializing the agent with prior memory consistently improves performance compared to the setting without initial memory, indicating that Mem2Evolve can effectively leverage accumulated experience to enhance the task execution performance. In [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

**Figure 5.** Figure 5: Specification of the tool for simulating the Piston Platform game. The specification includes the tool name and description, detailed definitions of input parameters and output formats—where each parameter is characterized by its name, type, description, and default value—as well as the core logic of the tool implementation, guiding subsequent tool creation [PITH_FULL_IMAGE:figures/full_fig_p018_5.png] view at source ↗

**Figure 6.** Figure 6: Specification of the probabilistic simulation expert. The specification defines the expert agent’s role, areas of expertise, suggested strategies or recommendations, and the list of tools available for use during task execution. exceeds a similarity threshold δ: a ∗ = arg max ai∈Magt n cos(q, hai ) | cos(q, hai ) > δo (12) If the set is empty, a new agent generation process is triggered. Tool Retrieval Sim… view at source ↗

**Figure 7.** Figure 7: Case Study 2 on YouTube Video Subtitle Extraction. When initialized with only a web search tool, (a) experience-centric frameworks fail to handle tasks situated beyond their capability boundary, such as retrieving internal video content, leading to incorrect answers based on general common sense. In contrast, (b) Mem2Evolve leverages the guidance of accumulated experience to dynamically generate high-quali… view at source ↗

**Figure 8.** Figure 8: Tool Experience: Using the GPT-4o for Image Analysis. This tool experience illustrates how to call the GPT-4o API to analyze images, where the agent can customize prompts to steer GPT-4o toward diverse and complex visual understanding tasks (e.g., recognition, counting, spatial reasoning, chart/diagram interpretation, and multimodal grounding). Each tool experience is organized into four fields: Title, Des… view at source ↗

**Figure 9.** Figure 9: Case Study 4: Experience-Guided Tool Generation for Attribute-Preserving Excel Parsing. This case illustrates how Experience Memory guides Mem2Evolve to generate task-appropriate tools that preserve critical non-textual attributes. When required to extract color-coded cells from an Excel file, Mem2Evolve leverages past experience to synthesize a tool capable of accurately retrieving both cell values and th… view at source ↗

read the original abstract

While large language model--powered agents can self-evolve by accumulating experience or by dynamically creating new assets (i.e., tools or expert agents), existing frameworks typically treat these two evolutionary processes in isolation. This separation overlooks their intrinsic interdependence: the former is inherently bounded by a manually predefined static toolset, while the latter generates new assets from scratch without experiential guidance, leading to limited capability growth and unstable evolution. To address this limitation, we introduce a novel paradigm of co-evolutionary Capability Expansion and Experience Distillation. Guided by this paradigm, we propose the \textbf{Mem$^{\textbf{2}}$Evolve}, which integrates two core components: \textbf{Experience Memory} and \textbf{Asset Memory}. Specifically, Mem$^{2}$Evolve leverages accumulated experience to guide the dynamic creation of assets, thereby expanding the agent's capability space while simultaneously acquiring new experience to achieve co-evolution. Extensive experiments across 6 task categories and 8 benchmarks demonstrate that Mem$^{2}$Evolve achieves improvement of 18.53\% over standard LLMs, 11.80\% over agents evolving solely through experience, and 6.46\% over those evolving solely through asset creation, establishing it as a substantially more effective and stable self-evolving agent framework. Code is available at: https://buaa-irip-llm.github.io/Mem2Evolve.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The joint memory loop is a clean incremental step for agent evolution, but the reported gains lack the ablations needed to confirm true co-evolution rather than additive complexity.

read the letter

The paper's core move is to wire experience memory and asset memory together so that distilled experience guides new tool or sub-agent creation, and the new assets in turn generate fresh experience. This closes the loop that prior single-route systems left open. They test it on six task categories and eight benchmarks, report gains of 18.53% over base LLMs, 11.80% over experience-only evolution, and 6.46% over asset-only evolution, and release the code. That combination of a joint architecture plus public implementation is the useful part.

Referee Report

2 major / 1 minor

Summary. The paper introduces Mem²Evolve, a self-evolving agent framework that integrates Experience Memory and Asset Memory under a co-evolutionary paradigm of capability expansion and experience distillation. It claims that isolating experience accumulation from dynamic asset creation leads to bounded or unstable growth, while their interdependence enables synergistic improvement. Experiments across 6 task categories and 8 benchmarks report gains of 18.53% over standard LLMs, 11.80% over experience-only evolution, and 6.46% over asset-creation-only evolution.

Significance. If the interdependence produces stable synergistic gains rather than additive or saturating effects, the framework could advance self-evolving agents beyond static toolsets or unguided creation. The availability of code at the provided link is a strength for reproducibility.

major comments (2)

[Experimental Evaluation] The central claim that co-evolution between Experience Memory and Asset Memory produces synergistic, stable growth (rather than interference or diminishing returns) is load-bearing for the reported deltas, yet the experimental evaluation provides no ablations that disable cross-guidance while holding total memory size and iteration count fixed, nor multi-iteration trajectories demonstrating continued improvement instead of plateau. The numerical gains could therefore arise from increased complexity rather than the asserted interdependence.
[Method and Experiments] The abstract and method description report percentage improvements without error bars, statistical significance tests, or details on how experience specifically guides asset creation (and vice versa). This leaves the implementation of the co-evolutionary guidance underspecified and the performance claims difficult to attribute to the proposed mechanism.

minor comments (1)

[Abstract] The abstract uses inconsistent formatting for Mem²Evolve (e.g., bold superscripts); ensure uniform notation and rendering across the manuscript.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback, which helps strengthen the validation of our co-evolutionary claims. We address each major comment below and have made revisions to the manuscript to incorporate additional controls, statistical reporting, and methodological details.

read point-by-point responses

Referee: [Experimental Evaluation] The central claim that co-evolution between Experience Memory and Asset Memory produces synergistic, stable growth (rather than interference or diminishing returns) is load-bearing for the reported deltas, yet the experimental evaluation provides no ablations that disable cross-guidance while holding total memory size and iteration count fixed, nor multi-iteration trajectories demonstrating continued improvement instead of plateau. The numerical gains could therefore arise from increased complexity rather than the asserted interdependence.

Authors: We agree that isolating the contribution of cross-guidance is essential. In the revised version, we add controlled ablations that disable experience-to-asset and asset-to-experience guidance while exactly matching total memory capacity and iteration budget; these show consistent drops of 5–9% relative to the full co-evolutionary setting, indicating the gains are not merely from added complexity. We also include multi-iteration performance curves (10 iterations) across the main benchmarks, demonstrating sustained improvement rather than early saturation. These results are reported in a new subsection of the experimental analysis. revision: yes
Referee: [Method and Experiments] The abstract and method description report percentage improvements without error bars, statistical significance tests, or details on how experience specifically guides asset creation (and vice versa). This leaves the implementation of the co-evolutionary guidance underspecified and the performance claims difficult to attribute to the proposed mechanism.

Authors: We have updated all tables and figures to include error bars (standard deviation over five independent runs) and paired t-test p-values for every reported comparison. The Method section has been expanded with explicit pseudocode for the two guidance directions, together with concrete prompt templates and examples showing how retrieved experiences condition asset generation and how newly created assets are used to produce distilled experiences. These additions make the co-evolutionary loop fully specified and reproducible from the text and released code. revision: yes

Circularity Check

0 steps flagged

No circularity: procedural framework with empirical results, no derivation or fitted predictions

full rationale

The paper introduces Mem²Evolve as a procedural integration of Experience Memory and Asset Memory under a co-evolutionary paradigm. Claims of stable interdependence and capability growth are asserted in the abstract and supported solely by benchmark comparisons (18.53%, 11.80%, 6.46% gains). No equations, first-principles derivations, parameter fits, or predictions appear in the provided text. Results are measured outcomes from experiments, not quantities forced by the framework definition itself. No self-citations, uniqueness theorems, or ansatzes are invoked to justify core components. The derivation chain is self-contained as an engineering design whose validity rests on external empirical tests rather than internal reduction.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

The central claim rests on the unproven premise that experience-guided asset creation and asset-generated experience form a stable positive feedback loop; no free parameters, axioms, or invented entities are enumerated in the abstract.

pith-pipeline@v0.9.0 · 5583 in / 1187 out tokens · 31637 ms · 2026-05-10T16:40:40.959216+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

45 extracted references · 45 canonical work pages

[1]

Dynamic LLM-Agent Network: An LLM-agent Collaboration Framework with Agent Team Optimization

RepoDebug: Repository-level multi-task and multi-language debugging evaluation of large lan- guage models. InFindings of the Association for Computational Linguistics: EMNLP 2025, pages 23784–23813, Suzhou, China. Association for Com- putational Linguistics. Zijun Liu, Yanzhe Zhang, Peng Li, Yang Liu, and Diyi Yang. 2023. Dynamic llm-agent network: An llm...

work page arXiv 2025
[2]

Each task description should be COMPLETE and SPECIFIC

work page
[3]

Use task numbers (1,2,3...) for dependencies

work page
[4]

dependencies

First task usually has no dependencies: "dependencies": []

work page
[5]

dependencies

If task 2 depends on task 1: "dependencies": [1]

work page
[6]

dependencies

If task 3 depends on tasks 1 and 2: "dependencies": [1, 2]

work page
[7]

Preserve the EXACT units and format requirements from the original query in task descriptions

work page
[8]

Why Code?

Output pure JSON format with no other content. Prompt for Assess Tool NeedBack to ToC You are a Principal Engineer. Your goal is to analyze the given multi-step task plan and determine the most efficient way to solve it. Your primary objective is to use LLM-Native capabilities and reuse existing tools. You will only determine that new tools are needed (ne...

work page
[9]

{tool_name}

name: string - professional general-purpose snake_case with domain prefix - Must be EXACTLY "{tool_name}" (do not rename or vary)

work page
[10]

description: string - DETAILED description (5-10 sentences) covering: - What the tool does (core functionality) - Key capabilities and features - Typical use cases and scenarios - Input/output data types and formats - Any limitations or constraints

work page
[11]

type": "object

input_schema: object - JSON Schema (Claude Desktop standard) defining parameters: {{ "type": "object", "properties": {{ "param_name": {{ "type": "string|number|boolean|array|object", "description": "Detailed parameter description", "enum": ["optional", "allowed", "values"], "default": "optional_default_value", "example": "input example" }} }}, "required":...

work page
[12]

type": "string|object

returns: object - Output format specification: {{ "type": "string|object", "description": "Detailed description of return value", "format": "json|text|structured", "schema": {{"optional": "output schema for structured returns"}} }}

work page
[13]

name": "{tool_name}

module_code: string - COMPLETE Python module source to be saved as ‘<tool_name>.py‘ Module requirements (STRICT): - Include necessary imports. - Define ONE public function implementing the tool with an EXPLICIT parameter list derived from input_schema (no **kwargs). - Implement robust validation and error handling per input_schema (types, required fields,...

work page
[14]

Review the outputs from previous agents to understand progress and context

work page
[15]

Utilize the available tools as appropriate

Analyze the task and decompose it if needed. Utilize the available tools as appropriate

work page
[16]

Define the current step you will complete, labeling it as ’CurrentStep’

work page
[17]

Choose one Action from the available tools to execute the current step. # Format example {format_example} Prompt for LLM as a JudgeBack to ToC You are a Judge LLM responsible for evaluating the entire execution trajectory of a task and determining whether it was completed correctly. OBJECTIVE Analyze the complete task execution trajectory and provide a co...

work page
[18]

Task Completion Assessment: - Does the final result correctly answer the original query? - Are all decomposed subtasks properly addressed? - Is the aggregated result logically coherent and complete?

work page
[19]

Agent Performance Assessment: - Did each agent execute its assigned subtask correctly? - Were the agents’ reasoning processes sound and effective? - Did agents properly utilize available tools?

work page
[20]

task_completed

Tool Effectiveness Assessment: - Did existing tools function correctly when used? - Were newly created tools (if any) implemented correctly? - Did tools produce expected outputs? EVALUATION OUTPUT FORMAT You must provide your evaluation in the following JSON format: {{ "task_completed": true/false, "completion_quality": "good/poor", "overall_assessment": ...

work page
[21]

Be objective and thorough in your evaluation

work page
[22]

Provide specific, actionable feedback

work page
[23]

Identify both strengths and weaknesses

work page
[24]

Focus on patterns that can inform future improvements

work page
[25]

Distinguish between agent failures and tool failures

work page
[26]

Consider the complexity of the task when evaluating performance

work page
[27]

How to

Ensure all JSON is properly formatted and valid Return ONLY the JSON evaluation, no additional text. Prompt for Tool Memory GenerationBack to ToC You are a technical documentation specialist creating comprehensive tool implementation guides. OBJECTIVE Generate a detailed markdown section that captures the implementation experience of this tool for future ...

work page
[28]

Write in clear, professional technical documentation style

work page
[29]

Use proper markdown formatting (## for main title, ### for subsections)

work page
[30]

Be concise but comprehensive

work page
[31]

Focus on reusability and understanding

work page
[32]

Include practical context, not just code

work page
[33]

Highlight key implementation patterns

work page
[34]

## How to

Note any important dependencies or requirements OUTPUT FORMAT Return ONLY the complete markdown section. Do not include any preamble, explanations, or commentary outside the document itself. The section MUST start with "## How to..." as the first line. Prompt for Success Agent Memory GenerationBack to ToC You are an experience synthesis specialist creatin...

work page
[35]

Write in clear, professional documentation style

work page
[37]

Be specific and actionable

work page
[38]

Focus on transferable knowledge

work page
[39]

Highlight decision-making processes

work page
[40]

Common Pitfalls When Parsing Inconsistent CSV Data Formats

Include both what worked and why it worked OUTPUT FORMAT Return ONLY the complete markdown memory entry. Do not include any preamble or explanations outside the memory entry itself. The entry MUST start with "##" as the first line. Prompt for Failure Agent Memory GenerationBack to ToC You are an error analysis specialist creating memory entries that help ...

work page
[41]

Write in clear, instructive style

work page
[42]

Use proper markdown formatting

work page
[43]

Be honest and specific about failures

work page
[44]

Provide actionable corrective guidance

work page
[45]

Focus on learning and improvement

work page
[46]

" " 16# Step 1: Validate inputs 17if not isinstance( num_balls ,int)ornum_balls < 1: 18raiseValueError ( f

Help prevent similar failures OUTPUT FORMAT Return ONLY the complete markdown memory entry. Do not include any preamble or explanations outside the memory entry itself. The entry MUST start with "##" as the first line. D Case Study D.1 Tool Implementation forSimulate Piston Platform GameBack to ToC As shown in Code 1, when performing complex probabilistic...

work page 2018